<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Andrey Belkin on Medium]]></title>
        <description><![CDATA[Stories by Andrey Belkin on Medium]]></description>
        <link>https://medium.com/@it.belkin?source=rss-f2a911e00a6b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*8us5PkTNjN8Ob-9K</url>
            <title>Stories by Andrey Belkin on Medium</title>
            <link>https://medium.com/@it.belkin?source=rss-f2a911e00a6b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 23:48:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@it.belkin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Artificial and biological neural networks]]></title>
            <link>https://medium.datadriveninvestor.com/artificial-and-biological-neural-networks-29cfb0a28008?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/29cfb0a28008</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[brain]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Sun, 08 Jan 2023 05:20:17 GMT</pubDate>
            <atom:updated>2023-01-08T13:28:26.662Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eVSZFZeAkVRobrzh" /></figure><p>This can be compared to the search for the Holy Grail, the search for the algorithm by which biological neural networks work. Of course, many will say that there is no grail, it’s all legends, and everything is already implemented in artificial neural networks, just wait for the development of this technology, computing resources and… and all the real artificial intelligence will be created. And, to understand the complex and confusing organ for this purpose is not necessary. But I hope there is a good portion of adventurers who will be interested in some considerations where to look for the Holy Grail. In this article we will analyze and compare the work of artificial neural networks with hypotheses about how biological neural networks work, and of course, accompany this with practical experiments, we will take apart a new artificial neural network, which is closer to its biological analogue in its principle of operation.</p><p><strong>So how does an artificial neural network work?</strong></p><p>We will start with the analysis of artificial neural networks. No, we will not describe the principles of networks here, numerous other sources will tell us about them better, here we will try to understand some fundamental essence of their work in terms of pattern recognition. One simple example demonstrates this very clearly:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tiKqDBdLXsagEs1N" /></figure><p>Imagine a network consisting of two parts of an encoder and a decoder, the input of which will be given examples from a set of handwritten digits MNIST and at the output, respectively, we duplicate the input. We will train it by the classical method, the back propagation of an error. The narrowest part of the network, the smallest layer located between the encoder and decoder will have only 2 neurons. After training, it will turn out that these 2 neurons will carry information in themselves, so to speak, encode images of a set of shapes of all numbers. If you place the activity of one neuron along the x axis, and the activity of another neuron along the y axis, then you can mark on the diagram which points correspond to each class, and use the decoder to reproduce the outline of the number corresponding to the points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/549/0*F8zzhxqVJzykor-S" /><figcaption><a href="https://n8python.github.io/mnistLatentSpace/">A great interactive example of this can be tried here</a></figcaption></figure><p>What we got is called the Small Representation (SR), “small” because we have only 2 real numbers reflecting a vector of 784 real numbers, and “representation” is because it reflects the external world inside the network, that is, knowing only these two values, we can estimate what the network “sees” at the moment, what figure and even the approximate form of its outline. We can say that the Small Representation (SR) is a kind of model of the world, only the network had to make a very simple model, small, just two numbers, although the external world is more complex and represents sets of 784 real numbers.</p><p>Let’s take a closer look at what this Small Representation (SR) is, we see several areas that will correspond to certain classes and neighboring points will correspond to very similarly shaped numbers. It turns out that SR can help greatly simplify the process of learning recognition. We can take a trained decoder, discard the encoder, add a small perceptron and quickly enough teach the new network to recognize handwritten digits (this is called transfer learning), that is, first our network, or its first configuration was trained only on images without connecting them to the actual number, such a self-training (without control) to get the SR, and then the next configuration using the obtained SR, is trained on matching images and digit labels, with control, and does it much faster than the network which.</p><p>In fact, here the class of figure is defined by some ranges of activity of the two neurons. These ranges or areas, as we see in the diagram, can be called some kind of embeddings, representations of classes within the network, which were formed in the process of self-learning.</p><p>Next we will talk about what is Small Representation (SR) in biological networks, in the brain, and how these embedding entities are represented in it, but first let us highlight some qualities of these entities for artificial neural networks. First, in SR the whole layer is involved, all neurons of the layer matter. Let’s imagine that we removed one of two neurons from the network, it would be problematic to detect exactly what the network “sees” by one neuron, it is not always possible to answer even the question, whether the network now sees, for example, the number “5”. Yes, there is pruning, but it is only possible when there are too many neurons, when we can remove neurons without consequences for SR. Secondly, the more neurons in a layer involved in SR, the harder it is to isolate fics, as each neuron creates its own dimension in the space of class representations, 2 neurons — 2 dimensions, 10 neurons — 10 dimensions… and all 10 will be interdependent. Thirdly, the obvious thing for artificial networks, all neurons of a layer or even the whole network, at each tact of training are involved, each neuron will get its correction of weights.</p><p>In contrast to classical artificial neural networks, biological networks reflect the neuron-detector approach. In a biological network, a neuron acts as a detector of a certain feature or group of features. One can say that individual neurons are responsible for features or embeddings here. This is confirmed by numerous experiments; there is even a joking name for this phenomenon — “granny neurons,” which suggests that any image in our brain will correspond to the activity of a certain cell; it is even possible to have such a cell for our granny’s image. Of course, when we say “neuron” in this case, we mean a whole group of active neurons. First, before the excitation reaches the notorious grandmother neuron, it passes through a number of other neurons; second, usually the activity of a whole group of neurons distributed throughout the cortex, a neuron ensemble, or a pattern of activity, is behind some kind of activity.</p><p>The formal neuron of artificial networks reflects quite well the principle of operation of a biological neuron, the weights of the formal neuron reflect the sizes of mediator portions released into the synaptic cleft during synaptic activation, and the cumulative effect of the neurotransmitter leading to the action potential is reflected in the threshold function of the formal neuron. That’s great, but</p><p>mapping of the primary visual cortex (“Eye, Brain, and Vision” by Hubel David, Torsten Wiesel) records neuronal activity to certain stimuli, for example, different orientations of lines, segments, and ends of segments. And the point here is not that cortical neurons work according to other principles, but rather in the very organization of these neurons.</p><p>In general, we can say that formal neurons of artificial networks handle information more efficiently than biological neurons, for example, as we have seen, it may take only two neurons to encode a set of handwritten digits, while encoding according to the neural-detector principle will require at least 10 neurons to at least encode 10 classes of digits, not to mention different forms of these digits. Then a reasonable question arises: why try to understand this principle at all, why search for this Grail?</p><p>Firstly, by working on the creation of a model of neural networks in which the neuron-detector approach will be applied, we can learn more about the principles on which biological neural networks work. Secondly, although the evolutionary process is not perfect and does not always lead to optimal solutions, evolution has been successful in saving energy and brain efficiency. With the neuron-detector approach there is no need to activate all neurons when training and running the model, yes, for a mathematical model this does not give much advantage in efficiency, but when implementing such a model in “iron” it can give significantly. Third, imagine how convenient it is to use a neural network consisting of neuron detectors, we can evaluate what the network is “thinking” at the moment by its neuron activity, we can understand how it makes decisions, we can manage these decisions by simply activating necessary neurons. Fourth, one-shot learning. If you have a neural detector responding to the word “table” and a neural detector responding to an image of a table, you would have no problem making a connection between them. Fifthly, it would be a more stable model in the matter of retraining, retraining (in a good sense), as well as physical destruction, if you damage the neuron-detector responsible for the number “5”, the remaining neurons allow good recognition of numbers, trivially because there is no close interaction of neuron activity level during detection. A few more arguments in favor of the prospects of such networks: this is the flexibility of their control in dynamics, which means the possibility to create high-level control algorithms for these networks, similar to the way emotional systems control memory and reactions in the nervous systems of animals and humans. It is also possible to evaluate novelty of incoming information at any abstract level at which the network operates, if our system has a detector that reacts to an external stimulus, and we can evaluate the degree of this reaction, and thus evaluate the proximity of the external stimulus to the internal memory of the system, which will be the evaluation of novelty. If novelty criteria become important for neural network control, this should bring us closer to understanding how to create real artificial intelligence, because novelty is a very important factor in human and animal behavior.</p><p><strong>How do biological neural networks encode information?</strong></p><p>The Small Representation (SR) for the brain is SDR (Sparse Distributed Representation). “Sparse” means that only a small number of neurons are active at one time, and “distributed” means that this activity is evenly distributed throughout the tissue. But the word “small” can also be applied to this term, since any representation in the brain will always be of a smaller dimension than the outside world. In fact, the sparseness is only the result of the neuro-detector approach, we will see in practice how it happens.</p><p><em>Many may associate the notion of SDR with the works of Jeff Hawkins, as he wrote extensively about SDR and applied it in his model of Hierarchical Temporal Memory or HTM, but SDR is a general experimentally observed phenomenon inherent in the nervous systems of all mammals (and insects) from mice to humans. The author does not share Jeff Hawkins views on HTM, because it is impossible for an HTM neuron to exist (as Jeff describes it), which does not fit into the conventional neuronal paradigm — again, in the author’s opinion.</em></p><p>Let’s try to create an artificial neural network based on the neuron-detector approach. We will be guided and inspired by the knowledge of the biological brain and the principles of its work.</p><p><strong>Network diagram and organization</strong></p><p>There is no tool that works equally well with normal and sparse data at the same time, so the mechanism of neuron detectors associated with external data and data arising inside the network will be different. The first layer of our model will receive ordinary input, in this case a 784 dimensional vector representing a digit from MNIST set, and we will get sparse (SDR) at the output, and all subsequent layers will receive SDR and output SDR. Essentially, the main task of the first layer is to convert the data into a sparse representation, there is no need to train this layer. After many variants to the principles of construction of this layer, I came to the simplest and best option, the use of several kernels of convolution manually selected. Sixteen kernels of size 5x5.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*623oJG4MDuQ5zjLn" /></figure><p>The choice of kernels is determined by the prevalence of patterns in the data we will be processing. For example, the figures from MNIST are characterized by empty, unfilled spaces, that is, in almost every example you can find a single 5x5 area where all pixels will be 0, such a kernel is reasonable to include in our set. Also the figures are made by lines of different thickness and orientation, which defines some more variations of kernels — lines of different orientation. Here it is important that the set turns out to be as diverse as possible.</p><p>As we know, the primary visual cortex is characterized by line-detector neurons of different orientation, each column actually tuned to a different stimulus. Hubel David and Torsten Wiesel in their work on mapping of the primary visual cortex introduced the term — hypercolumn or module, which is a set of all the columns that have a common receptive field, for example, a small area on the screen where luminous lines of different orientation are projected. That is, as a part of the module (the term module is more acceptable to me than hypercolumn, as there is no consonance with the term column, so hereinafter I will apply this name) there are all representations of orientations of lines for both eyes and it makes about 1000 columns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/830/0*bkyOdTJ0qA4-dwiQ" /><figcaption><em>Hypercolumn or Module, illustrations from The Eye, the Brain, and Sight by Hubel David</em></figcaption></figure><p>In our variant, there will be only 16 neuron detectors in a module (yes, our neuron detector is equivalent to one column in the cortex), each corresponding to its own kernel. Each module will be responsible for its own fragment of the 5x5 picture, and, like in the cortex, the receptive fields of the columns usually overlap, so we will divide the total picture into 5x5 fragments in increments of 1.</p><p>The modular principle of organization is fundamental to our network, both for the first layer and for subsequent layers. The main feature of modular construction is that there is always one winner in a module. That is, there is a competition among neuron-detectors of the module, which will determine which neuron in the module will have priority in activation and learning. Just as living columns have lateral inhibition, due to which columns with weak activity are completely suppressed, and due to the effect of irradiation (aspirations of excitement will spread) the winning columns are strengthened in their activity.</p><p>In my previous works, I described experiments with a pixel shader in which lateral inhibition and irradiation, which shows that any initial activity is converted into small point-like distributed foci. This is the reason for the distributed and sparse activity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/0*bFGUsHGQjOIpXH5X" /></figure><p>The cortex has no explicit module boundaries, but when creating a mathematical model of neural networks we can make it more precise and structured, to some extent this is to our advantage. Therefore, lateral inhibition can be replaced by the rule: only one neuron-detector can be active in a module.</p><p>Initial activity of each neuron in the first layer is estimated by Manhattan distance, but eventually, in each module we will consider only one active, the one that will have maximal closeness to the input vector in the module. The module operation may remind us of the principles of the Kohonen network, our future model is as if composed of many small Kohonen networks.</p><p>Examples of work of the first layer. The input vector is 28x28, the core is 5x5 step 1, there are 16 neuron detectors in each module, so the output is 24x24x16 which is converted into planar form 96x96. This output of the first layer is an SDR for any input vector.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/0*08TQAmqeWH6DfKd7" /></figure><p>Planarity or the representation of data in two-dimensional form in our case is very important, because the subsequent layers will receive data in this form, it is even more visual and similar to the way the data in the brain cortex is represented.</p><p>As it was mentioned before, the main task of the first layer is to represent external stimuli in the form of SDR. Since the number of neurons in the module is 16, the ratio of ones to zeros in the resulting SDR will always be 1 to 15, for any input vector. Our first layer is not trained, it only transforms data.</p><p>Since the layer performs the usual data transformation, it is possible to obtain an inverse transformation, which shows that the data transformation occurs with some loss of quality, but this is a necessary sacrifice, because the obtained data will be absolutely normalized, the total sum of the obtained vector will always be the same, regardless of what is fed to the network input. In biology, similar transformations take place already at the level of the retina, thanks to the on-, off-cells, no matter what is projected onto the retina the total nerve activity coming from the eye will be constant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/576/0*yIiwpK0HQ1x6-UWJ" /></figure><p>Neurons of subsequent layers will receive sparse vectors as input. If neurons of the first layer, we can say, extracted features from the input data (sticks, dots, empties), then neurons of subsequent layers will generalize, i.e. generalize the extracted features.</p><p><strong>Unsupervised bottom-up learning</strong></p><p>Self-learning is the main clue as to where to look for our Holy Grail. It can be said that usually learning in animals and humans is mostly based on data that are not labeled, often this data is incomplete and noisy. And labels appear in the process of learning only sometimes, in the form of comparison of modalities, that is, learning by labeling, supervised learning is only auxiliary, because the basis is self-learning.</p><p>If we talk about the problem of pattern recognition, in essence we need an algorithm that would allow us to form a diverse set of neuron detectors, which would respond to explicitly expressed representatives of their classes. Earlier in my research I used the Kohonen network algorithm and its variations, but it does not work well with sparse representation, which did not allow me to create a multilayer version of the network. And the neuron of this network determines its activity by Manhattan distance, but I was inclined to a simpler version of the neuron, as a simple adder.</p><p>And solving the classification problem on simple adders turned out to be very easy. I was interested in <a href="https://www.frontiersin.org/articles/10.3389/fncom.2015.00099/full">the article</a>, which told about successful clustering of MNIST using spike networks, <a href="https://github.com/peter-u-diehl/stdp-mnist">the solution</a> was very big, cumbersome and complicated for me.</p><p>So I tried to implement the mechanics of this clustering without anything unnecessary, the result fits literally in a few lines of code (<a href="https://colab.research.google.com/drive/1Mjx-ULP8oQKCw_Pf6uMS9s7-qIbbtAHZ?usp=sharing">link to Google colab</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/564/0*1z7xwxlNmW0-FXTz" /><figcaption><em>Network weights after 10,000 MNIST examples</em></figcaption></figure><p>The example above weighs 256 neurons in the layer (16x16). Only one, the most active neuron in the whole layer can be trained per cycle of the network.</p><p>The training follows a very simple formula:</p><pre>map[:, winner] += rate * (inp - 0.5)</pre><p>to the winner’s weights we add the product of the coefficient of the training rate (rate=0.2) and the difference of the input minus 0.5. Why exactly 0.5? — It is a trivial average of the input range, it is a real number from 0 to 1. Let’s call this number — the target intensity. The target intensity allows you to draw a line between weak and strong input. This formula essentially weakens weak inputs and strengthens strong ones. And the weaker the input, the stronger we decrease the weight associated with it and vice versa. Of course, due to this, not all data is suitable for this algorithm, where the number in the vector does not mean intensity only, but carries more complex information, most likely it is better to use Kohonen nets, but for MNIST it works well enough. Also this algorithm has a distinct advantage, there is practically no “super winner” or leader problem. The essence of it is that one neuron that wins most often takes all the learning. Thanks to the training of neighbors in Kohonen networks, this problem does not manifest itself explicitly. The thing is that if the starting sum of weights is big enough when using the above algorithm, then the more a neuron is trained, the lower is its activation level per input vector, compared to less trained neurons, thus less trained neurons will have more chances to win. This is of key importance at the beginning of learning, when a neuron is just acquiring its specialization. So it was logical to make the value of all weights at network initialization equal to one.</p><p>At birth the brain has the maximum number of neurons and the first days and weeks the maximum number of synaptic connections are unfolded. Then, as the brain grows and learns, there is a general degradation and loss of synaptic connections and loss of nerve cells. This process is called pruning and is particularly large for the human brain by the first three years after birth. Pruning allows neurons to specialize more precisely, and this can be compared to what happens in our clustering algorithm; initially all weights are equal to 1, then in the learning process the extra is reduced to zero.</p><p>But this algorithm has a drawback, it strives for the smallest sum of weights, with more and more learning the activation level will decrease, the sum of weights will also decrease, and to win the neurons will specialize on examples with less and less sum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/564/0*jGaNWTIu_aNRWKLN" /><figcaption><em>Network weights after 10 epochs on all 60,000 examples</em></figcaption></figure><p>To eliminate this problem, the solution is to normalize the input vector, and to bring the sum of the input vector to a constant. And the best solution for this is SDR.</p><p>But in fact, the simple combination of this algorithm and SDR does not solve the problem of reducing the total sum of weights, although the diversity of classes will be preserved. It is necessary to bring even more stability to the algorithm.</p><p>The general idea of a biological neuron is that it strengthens in regularly repeated connections, while rarely used connections are degraded and lost, some stable pathways of nerve signals propagation are formed. To implement this, we need to keep some kind of statistics for each synapse, how often and how effectively it is used. Therefore, in addition to the weight map, we will have a “temperature trace” or trace map, the value of this map will be in the range from 0 to 1. Every time both neurons are active, the synapse connecting them will add +0.3 to trace, but during other cycles of the network trace will gradually decrease by -0.001. Thus synapses that will be regularly used effectively will have a high trace value close to 1. In contrast, synapses that are hardly used will have a trace value of zero or close to zero. This indicator will have a direct effect on the value of the synapse weight, because frequently used synapses should be strengthened. The final formula for the resulting algorithm is as follows:</p><pre>map[winner,:] = (1 - rate)*map[winner,:] + rate*(inp - 0.5) + trace[winner,:]*rate</pre><p>In addition to the well-known formula, here the neuron weights increase by the product of trace and rate, thus the frequently used synapse strengthens, in contrast to this there is regular degradation of synapses, their weights decrease modulo (1 — rate)*map[winner,:]. Since constant growth in biological cells is impossible, there is always some limited resource, so there is heterosynaptic competition. Also the weights are limited in the range from -1 to 1, where -1 is the maximum strong inhibitory synapse and 1 is the maximum strong inducing synapse.</p><p>The described algorithm is quite simple and allows to perform self-learning of the network even organized in many hierarchical layers.</p><p>Next, let’s look at a particular configuration of a multilayer network as an example.</p><p>For the considered variant of the network the modules of the next layer will also contain 16 neuron-detectors each, which will have a common receptive field. The receptive field of a neuron will have the size 20x20, which is 25 (5x5) modules of the previous layer. Receptive fields of modules will overlap in increments of 4. The next layer will contain 400 (20x20) modules of 16 neurons, totaling 6,400 (80x80) neurons. Let me remind you that the first layer is 96x96 (9,216) neurons.</p><p>Using a similar scheme, let’s add two more layers, the first of which will have a module size of 16 (4x4) neurons. The third layer — the size of the receptive field will be 20x20, step 4, total 64x64 neurons. The fourth layer — receptive field will cover all previous layers 64x64, but we will make the module size 36, actually, the activity of this layer will be characterized by only one neuron out of 36, to get a clear pronounced detector of one digit. And the last fifth, the output layer will contain only 10 neurons by the number of recognized classes. This layer is needed to match the result of the previous layers and the digit labels. This layer will have only one module, so only one neuron from this layer can be active. Each neuron of the output layer will have the same receptive field of size 6x6 (=36).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*scyevZV36uNU_IXS" /></figure><p>The result is the scheme: 28x28 (receptors) → 96x96 (the first layer, untrained, simply converts the input signal to SDR) → 80x80 (second layer) → 64x64 (third layer) → 6x6 (fourth layer) → 10 (output layer). The multilayer of this network was chosen to demonstrate all aspects of its operation, in particular “deep learning”, for the MNIST task a simpler configuration can also be used.</p><p>The output layer has only 10 neurons by the number of recognized classes; all neurons of the output layer are included into one module. This means that all neurons of the output layer have a common receptive field equal to the size of the previous layer (6x6 in this example), and only one neuron of the output layer can be active.</p><p>The penultimate layer consists of one module, it will form a certain representation during self-training, each of 36 neurons will respond to a selected type of input signal, while training of the last layer is based on matching the obtained representation with the label of the class represented by the example network. If the label coincides, then we increase the synapse weight by the rate (0.1), and for all other synapses that are connected with other classes, the weight is reduced by the same value only divided by the total number of classes, i.e. by a number ten times smaller.</p><p>As a result, self-training on 5,000 examples presented to the network once, the quality is about <strong>77%</strong>. That is, we clearly see that self-learning works and there is an allocation of digit classes, even without reference to the label. But we can improve this result by spreading the influence of the class label deeper into the layers.</p><p><strong>Supervised learning or backpreference</strong></p><p>The backpropagation algorithm for connective neural networks has become an incredibly powerful engine of machine learning, and it has enabled artificial neural networks to produce amazing results. But this algorithm does not match what we know about the biological nervous system. And it’s not a question of spreading signals back through the network, it’s a question of transmitting a lot of information, transmitting an error value, and there are no known mechanisms, mechanics or organelles in biology that would even evaluate this error itself. There are many articles that propose supposedly biologically close variants of backpropagation, but all of this, if assessed from the biology side, is implausible.</p><p>On the one hand, there are feedback connections between different cortical areas involved in hierarchical information processing, as analogs of layers in artificial neural networks, and they sometimes number several times more than direct connections. There is a constant circulation of excitation between cortex areas — reverberation. But there are no complicated calculations. Calculations are made on the organelle-membrane, and the algorithm of these calculations is defined by the Hodgkin-Huxley model, that’s all, in this system of knowledge there is no place for backprop and error calculations.</p><p>And yet a neuron at a low hierarchical level with a limited receptive field, no matter how good its self-learning algorithm is, cannot learn effectively without a high-level conductor, without cues coming from a common task or goal.</p><p>The modular organization of the network and SDR allows us to apply the backpreference algorithm. Neurons in the module compete with each other, since there can be only one winner in the module, which neuron wins is determined by ascending connections and their weights, but it is possible to give some neurons a preference in this competition thanks to descending connections. It is possible to take any neuron of the last layer and evaluate the activity of which neurons of the previous layer will most likely lead this neuron to the victory, and then it is possible to go down to the lower levels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/966/0*sYyOPI-SNWk3EJ5w" /></figure><p>Knowing the class label, we can select neurons of the last layer that are most relevant to this label, then moving down from layer to layer, designate those neurons that will eventually give a preference to the right neurons of the last layer, then give those neurons an advantage in winning by adding some value to the final sum of their activations. So we only slightly interfere with the process of self-learning, directing it in a more appropriate direction. We can adjust how much supervised learning will influence unsupervised learning. It makes sense if supervised learning is stronger on the last layers, and self-training is stronger on the first layers, which are closer to the receptor input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Y6kY6DE1EvulSQWA" /></figure><p>In our example there are three trainable layers, the last, third of them supervised learning operates with a coefficient of 1, which says 100% will win the neuron that gets the preferences. For the second layer the coefficient is 0.5, and on the first layer only 0.1, here self-training has a greater advantage.</p><p>It is a very flexible combination of supervised and unsupervised learning, you can turn off control completely at any time and let self-learning take control, and when the necessary labels appear add in control learning. It’s very similar to how the biological brain works and learns.</p><p>At the beginning of learning, when the output layer has not yet determined the correlation of the last layer neurons and class labels, supervised learning does not occur at all, only weak correlations are formed through self-training, which are then picked up by supervised self-training and strengthened.</p><p>The result using supervised learning on 5,000 examples presented to the network once, the quality is about<strong> 93%</strong>. Of course the algorithm is inferior in quality metrics to the connection of neural networks with backprop, but it is a very decent result, given that the algorithm does not have any methods for convergence and error calculation. Increasing the dimensionality of the network, and hence its potential capacity with the same training parameters can slightly improve the quality result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zz48mCLA9-swl-z7" /></figure><p><strong>Interpretability of artificial neural network based on the neuro-detector approach</strong></p><p>The question of the interpretability of the network is very important, it is not enough to take a model to train it on a large amount of data and trust it completely, especially when it comes to the life and health of people. It is necessary to understand well why this model makes certain decisions, it must be possible to predict its actions on any external stimulus. In this, neural networks based on the neuro-detector approach will be unrivaled.</p><p>We have already touched upon the topic of interpretability speaking about small representations, in the example where the latent space has only two dimensions as in the case of two neurons between encoder and decoder, in this example we can easily interpret the learning results, but only one of the narrowest layers. But, for example, in face recognition tasks, the latent space is quite large and has, for example, 18 parameters. It will be quite difficult to represent in a clear and readable form the 18-dimensional dimension and to compare it with face images, not to mention the other layers of the network.</p><p>For networks based on neuro-detector approach it is easy to interpret the results, it is possible to perform inverse transformations and get values of receptive field activity that will most likely lead to activity of a particular neuron, and this can be done for any neuron of any layer. In fact, the inverse influence propagation algorithm performs inverse transformations, we will just apply it to interpret the training results.</p><p>The ability to interpret the results of network training is needed not only to increase the confidence of the network, but also to be able to exclude unacceptable actions in its work by correcting the network, directly changing its weights, blocking the work of specific neurons, rather than revise the contents of the dataset or apply additional algorithms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/588/0*m5mMhBcyZ_lpKkQW" /><figcaption><em>Reverse analysis by output layer (3rd learner)</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/0*zd0jAmoLA5XEqy2m" /><figcaption><em>Reverse analysis by hidden layer (2nd learner)</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/586/0*j39EiK4pJwB1VGPq" /><figcaption><em>Reverse analysis by hidden layer (1st learner)</em></figcaption></figure><p><strong>Local minimum of connection networks with backprop</strong></p><p>Artificial neural networks and in particular the error back propagation algorithm is one of Man’s greatest discoveries, which will still contribute much to progress. This perfect mathematical tool can transform data into a function, but in the matter of creating general artificial intelligence or human-level intelligence, the development of networks as a mathematical tool will not lead to the desired result. It makes no sense to strive to surpass backprop in quality metrics. The study of neurobiology, neuropsychology and physiology of the nervous system provides insight into the principles on which cognitive functions are based, but these principles are in no way consistent with the mechanisms of classical artificial networks. Therefore, I have long been searching for the “Holy Grail” of a neural network algorithm corresponding to the paradigm of biological brain network operation, and the next stage of research will focus on application of knowledge of network control based on the neuro-detector approach using emotional mechanisms and implementation of learning mechanisms with reinforcement based on this network.</p><p>If we imagine the task of creating general-purpose artificial intelligence as a space of solutions, then with connective networks we are at a kind of local minimum, it is impossible not to note the greatest achievements obtained by models based on this algorithm. But there is an alternative, which is only nascent, which does not yet have bright and significant results, and in the future, who knows, with the development will give an opportunity to solve the problem of creating artificial intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/0*kfnzcaJLU4MAAo82" /></figure><p><a href="https://colab.research.google.com/drive/1WikVu0weaTE9XAWnb8fUnfn91YyTSljV?usp=sharing">Source on Google colab</a></p><p>Subscribe to DDIntel <a href="https://ddintel.datadriveninvestor.com/">Here</a>.</p><p>Visit our website here: <a href="https://www.datadriveninvestor.com/">https://www.datadriveninvestor.com</a></p><p>Join our network here: <a href="https://datadriveninvestor.com/collaborate">https://datadriveninvestor.com/collaborate</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29cfb0a28008" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/artificial-and-biological-neural-networks-29cfb0a28008">Artificial and biological neural networks</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Living creature learning and the computer neural network]]></title>
            <link>https://medium.datadriveninvestor.com/living-creature-learning-and-the-computer-neural-network-597e1ba2fc59?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/597e1ba2fc59</guid>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neurons]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Mon, 23 Aug 2021 11:38:05 GMT</pubDate>
            <atom:updated>2021-08-23T11:38:05.824Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s look at how the living brain learns. To what extent its learning is similar or not similar to the way machines do it. Let’s try to model some aspects of learning.</p><p>The terms unsupervised and supervised learning have taken root in machine learning. Unsupervised learning is learning from unlabeled data or examples. And supervised learning is usually training on some marked-up data, training on examples when the result is regulated and corrected by some external mechanism taking into account this very markup. Sometimes the term unsupervised learning is used when we have some agent, which we put in some environment, and the agent initially does not know by what rules and laws the environment operate, and without external help, the agent learns to interact with this environment. If the agent has some mechanism for evaluating the achievement of the goal, it can already be called the term — learning with reinforcement.</p><p>How correct and applicable are these terms to the training of living organisms?</p><p>No creatures in nature have a “clean” nervous system at birth and they do not have to adapt to their environment using some universal learning algorithm from scratch. The nervous system of the simplest animals already contains all reflexes and instructions necessary for life in their traditional environment. Their ability to adapt to changes in external conditions is very limited.</p><p>With the evolutionary development of living beings on Earth and the emergence of the cerebral cortex in the nervous system, the possibilities of learning and adaptation have greatly increased. But these learning opportunities also rely on innate skills and mechanisms.</p><p>The more developed the cerebral cortex is, the greater the role of learning in the animal’s life activity. For example, a newborn calf is on its feet within a few hours of birth and can walk, though uncertainly, and this knowledge is given to it at birth. Man holds the record for the longest period of time in which he remains completely non-sufficient and dependent on other individuals at birth. But in his further development Man overtakes all other animals.</p><p>But also the nervous system of Man is not a “blank slate”, there is no need to learn: to breathe, to drink milk, to interact with the mother through smiling and crying, there are many reflexes. Further human learning itself will take place according to a strictly defined program also laid down from birth; there are a number of mechanisms that will guide this learning in a certain direction.</p><p>I’m not saying that a universal algorithm for learning from a “blank slate” is impossible, just that there are no examples of it in the living nature. So to look for such an algorithm is a very difficult task, it is almost a search for something in total darkness. But we have examples of animal intelligence and human intelligence, understanding their nature is an easier way. If you understand how human intelligence works, understand how it learns, you can get more universal algorithms.</p><p>It seems that human learning has a continuous character, has no clearly marked out stages, its learning is universal, it learns both from unmarked data and from examples of comparison of different modalities. For the nervous system, it practically does not matter if a person learns, for example, by reading a book, and the remarks made by the teacher can also be referred to simply information similar to that received from the book.</p><p>Two types of learning can be distinguished in the psychology and physiology of living things: reinforcement learning and statistical learning.</p><p>With reinforcement learning, everything is quite clear — this is when an action, event, or stimulus is combined with factors important to the organism, thereby fixing the action or linking the event or stimulus to an important need for the organism. An example would be Pavlov’s experiments with dogs. The term reinforcement learning is now actively migrating into the field of machine learning as well.</p><p>So what is statistical learning for living things? This is a very important aspect of learning that should also be borrowed by machines. Imagine a newborn baby, it has many basic reflexes, it has small and chaotic reflex motor activity, its brain is almost fully formed, it has almost a complete set of neurons, even more, complete than an adult. And from the very first moments of his life, his learning has already begun. His brain begins to fill with an almost continuous stream of information in the form of “video and audio data,” tactile sensations, information from muscles, vestibular organs, and more. So what does the brain do with all this information? What does it learn? After all, for the most part, the information that we have to work with is not marked up, it’s just a continuous flow of information. And the brain tries to break this continuity by forming anchors. An anchor is a term in psychology for some kind of image mark or associative link in a person’s mind, which is responsible for evoking a particular state in him. The brain creates marks in the flow of information, and it creates them on the basis of repetition.</p><p>Statistical learning shows up well in language learning and the ability to distinguish sounds. For example, in experiments where children 10–12 months old need to distinguish among the sounds “ra… ra…” sometimes appearing sound “la”, among English-speaking children and children from Japan there are big differences. The fact that in English the [r] and [l] sounds are different, they are two different sounds. In Japanese, there is no difference between these two sounds. That is why Japanese children have a hard time grasping this difference, unlike English-speaking children. But at 6–8 months of age, all children pass these tests to the same average degree.</p><p>That is, our brain selects certain patterns from the general flow of information, and as we see in different cultures or we can say with a different history of the information received, different labels can be distinguished.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*NBC-v3UUhkVw3U_Rx0Gp2A.jpeg" /></figure><p>This illustration demonstrates that the brain can form anchors in a continuous spectrum of color and that different people’s brains can do this in different ways. It should be understood that we generally perceive the world only through the prism of formed anchors, and we cease to feel the continuity of the flow of information.</p><p>Through statistical learning, our brain forms a kind of model of the world around us, its ontology, the tops of which will be the anchors. Having such an ontology, further learning takes place on its basis, which greatly accelerates learning. We perceive the world only through the prism of the internal ontology of the world and its vertices.</p><p>There are certain stages of development and learning in living beings with a developed nervous system. The first weeks after birth can be key in the development of the visual system, if the brain is not given access to visual information during this period, it will lead to irreversible consequences, and even if the brain is given access to visual information at a later period, it is not possible to fully train the brain to “see”. This does not happen in adulthood, and loss of vision for years does not lead to loss of the brain’s ability to “see,” after the restoration of vision, for example, with cataracts. These periods are pronounced for many aspects of learning. Newborn beings have the highest adaptive capacities, and in the shortest period of time, adaptation to the external conditions of the environment and the capabilities of the organism itself takes place. What if the organism appeared in an environment where there is a total absence of light or the newborn organism will have a defect in the visual system, it is worth adapting to this quickly and redirecting all the resources of the organism to what works, to what is relevant and necessary. In the first moments of life, the brain sort of sorbs the type and nature of the information received from the external environment, in which it will exist and further learning.</p><p>For example, experiments on kittens in their first weeks, who are placed in an environment where only vertical lines are present, where no horizontal lines existed. In the world of newborn kittens, no horizontal lines existed, this made them immune to them. The instinctive fear of snakes did not work, if a model snake lay across the kitten’s view, he did not actually see the snake, unlike the same kittens who were lucky enough to develop in a normal environment.</p><h3>Teaching a fly</h3><p>Don’t be surprised, but to get closer to understanding the learning of living creatures with a developed nervous system, we need to understand how a fly learns. The fact is that the fruit fly turned out to be a very convenient model animal, and several scientific publications appeared that describe in detail the structure of its nervous system responsible for the sense of smell. In particular, an article with the resounding title <a href="https://www.pnas.org/content/pnas/112/30/9460.full.pdf">“What the Fly’s Nose Tells the Fly’s Brain” by Charles F. Stevens</a>.</p><p>Based on such parses of the fruit fly nervous system, some machine learning researchers began to develop their models (<a href="https://medium.com/@jaiyamsharma/efficient-nearest-neighbors-inspired-by-the-fruit-fly-brain-6ef8fed416ee">1</a>, <a href="https://science.sciencemag.org/content/358/6364/793">2</a>) and, in particular, create <a href="https://flyvec.org/?neuronIndex=0&amp;queryPhrase=Senate%2520majority%2520leader%2520discussed%2520the%2520issue%2520with%2520the%2520members%2520of%2520the%2520committee&amp;showQueryResults=true">language models</a>.</p><p>Let’s take a brief look at how the fly’s olfactory analyzer is organized.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*sjCdjj5zj6C8PJfFMc7gLA.png" /><figcaption>Fly’s neural network</figcaption></figure><p>The inputs to the network are olfactory receptors (X). The fruit fly has approximately 50 olfactory receptors, which respond to the concentration of certain molecules in the air; accordingly, each receptor responds to a different type of molecule or chemical marker. Odors can consist of many chemical markers like a spectral slice, each odor is characterized by a different pattern of receptor activity. The next layer of receptor duplicating neurons (A), also 50 in number, does a little signal transformation, essentially repeating the signals from the receptors. Then comes the Kenyon cells (B) — a ganglion with about 2 000 neurons, the connections with the previous layer are evenly distributed so that each neuron from ganglion (B) has a connection with about 10% (6 out of 50) neurons of the input layer. The activity of this ganglion will be suppressed by the reverse inhibitory system © so that only 5% of neurons at most of the whole ganglion (100 out of 2,000) are inactivity. Thus, from a vector of dimension 50, we get a vector of dimension 2 000, but this vector is sparse, i.e. it has a large number of zeros (95%). Many researchers have found in this an association with a hash function, in my opinion, this comparison is incorrect and even harmful, as it can mislead, I will explain in detail why I think so.</p><p>The above-described part of the nervous system of the fruit fly we will conventionally call the encoder. Nerve paths from the encoder ganglion go on to another cluster of neurons, and this nerve node is responsible for how the fly will react to the smell it catches. We will call this area the decoder. So far, there is no unequivocal understanding of whether changes in synapses occur as a manifestation of learning between layers A and B; most likely, these connections are formed during the development of the fly’s organism and do not change during life. But unequivocally, it is possible to assert learning and, consequently, change the weights of synapses in the decoder, and most likely here, Hebb’s rule will work.</p><p>One can think that the fruit fly encoder is a kind of transformer of a 50-dimensional real smell-detecting vector into 2 000-dimensional binary sparse code, hash-code. You can even estimate the size of this transformation: 2 000*1999*1998*…*1901≈1*10**329 (100 of 2 000 active). Yes, with such capabilities a fly can enjoy countless shades of scents of flowers, herbs, and fruits in its short life. But let’s evaluate what classes the fly needs to allocate odors to in order to successfully survive and leave offspring. I was only able to identify 4 classes (perhaps fly experts can add more):</p><ul><li>an attractive nutritious, sweet, or putrid odor of fruit, indicating a food source as well as places where larvae can be deposited (fruit flies lay larvae in the fruit of plants);</li><li>repulsive, dangerous smells of poisons, toxins, or smoke. The unambiguous reaction to these smells is to flee, to avoid;</li><li>attraction is associated with reproduction. Love and reproduction in flies are regulated by odors, and it is important to distinguish breeding attraction from feeding attraction;</li><li>neutral odors are a very important class of odors. There will be a lot of them in a fly’s life and it is advisable not to react to them in any way, otherwise, there will be no time to eat and reproduce.</li></ul><p>Four in all. Four forms of behavior. Is it reasonable to use a hash code that can encode 1*10**329 values for a total of 4 classes?</p><p>As Charles F. Stevens points out in his article, with this encoder configuration, two random smells are highly likely to have two non-intersecting sets of active Kenyon cells, and two very close smells will have such a similar pattern of activity of these cells that it will not make it possible to distinguish between the smells. It turns out that the possibility of encoding odors by the mushroom body of the fly was initially overestimated by us.</p><p>Another important fact is that the connections between neurons in the mushroom body differ strikingly between different individuals. Apparently, there is no information in the fly genome about which neurons should form synapses between which neurons. The genome stores only some general scheme of the nervous system, the general arrangement of ganglia, the density of connections, and some other parameters. During the development of the organism and nervous system, depending on the location of the neuron, it will grow its offshoots, in certain directions guided by chemical gradients, and form synapses with other neurons at the distances specified by the genome. Various mechanisms like competitive innervation will also work, which will limit the formation of multiple synapses with the same neuron, due to which a more even distribution of connections will be formed. In other words, high precision of organization of connections is not required for this scheme to work, and it is also possible due to some redundancy in the number of neurons.</p><p>When reverse-engineering such biological neural circuits, it is often forgotten that we are not dealing with mathematical models of neurons, but with biological cells. And these biological cells do not work a priori as precise automata, their accuracy is influenced by many external and internal factors, which is compensated by redundancy in their number. Neurons form the nervous tissue, and it is the properties of the nervous tissue that determine the nature of information processing by it. In this case, the activity of only 5% of the entire population of Kenyon cells suggests that we can accurately and clearly distinguish only 20 smells (100 / 5) and 100 neurons are allocated for each such smell. Where a biological neural network uses 2,000 neurons everything can be reduced to 20 neurons mathematically. If a ratio of two orders of magnitude seems large to you, then with mammals the figure would be three orders of magnitude. I wrote more about this in my article “<a href="https://blog.goodaudience.com/comparison-of-the-brain-with-a-computer-neural-network-3ecd3c1fef57">Comparison of the brain with a computer neural network</a>”. You often see how you multiply the total number of neurons and the average number of synapses per cell to estimate the size of human brain memory, the resulting number doesn’t really reflect anything.</p><p>So, we are dealing not with a hash code, but with groups of neuron detectors, which number about 20, which is comparable to 4 selected classes. To better understand how all this works, let us represent the solution space of this network in a simplified two-dimensional version. Each point of our two-dimensional solution space will represent a particular smell. Then this space can be divided into areas corresponding to the detected odors, which will be similar to the <a href="https://en.wikipedia.org/wiki/Voronoi_diagram">Voronoi diagram</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/202/1*z0yz6AUJtpTd5tJFrmM8wQ.png" /></figure><p>The dots in the image show the reference vectors, which are the odors that our neuron-detector groups are most attuned to. The colored areas around the reference vectors are loci, areas of similar odors that will not differ when detected. Like the Voronoi diagram, the mushroom body breaks down the entire imaginary space of smells into arbitrary areas, and, accordingly, these areas will be formed differently in different individuals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/632/1*6AQPrKKhj3tQo_jLtY_VxA.png" /></figure><p>Although some reference odor vectors must be “reserved” at the DNA level, at least one odor for each of the 3 classes. This is required by the reflex theory, which says that the formation of new conditioned reflexes is possible only on the basis of already existing reflexes. It is hard to imagine that nature did not take care that the newborn fly did not have to determine empirically what in this world is edible and what is poisonous, and what is female/male of its species. For this purpose, it is enough to reserve only three reference vectors of odors (no need for neutral odors), the rest of the odors can be dealt with in the process of life activity and learning.</p><p>This solution for an odor analyzer greatly speeds up and simplifies the training of the fly, if necessary, because it is enough to assign a class to the locus for training, thus adding a class to the whole array of similar odors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/204/1*4VffIaRhGwDvTwbndOkHAw.png" /></figure><p>But this system has disadvantages. What if the odors we need to distinguish are at the same locus? In this case, it is not possible to train flies to distinguish them, but due to the fact that most of the reference odor vectors of different individuals are randomly placed, there are bound to be individuals in the population whose two odors need to be distinguished are at different loci. In this case, selective selection may breed flies that can distinguish any odor by a certain placement of the reference odor vectors, or by changing certain receptors. A good move on the part of nature by allowing some variation in the detection of odors for individuals, then making it possible to cover the entire spectrum of odors for the population as a whole.</p><p>Let’s take a look at one variant of this development, which is only available to the fruit fly in one mutation. <a href="https://en.wikipedia.org/wiki/Gene_duplication">Duplication</a> is a type of chromosomal mutation in which a section of the chromosome turns out to be doubled. Very often, duplication means an increase in the volume of some tissues, and sometimes even the doubling of an organ. Imagine a fruit fly in which, as a result of such a mutation, the ganglion tissue with Kenyon cells doubled from 2,000 to 4,000 cells; this is a fairly realistic scenario. On the one hand, such an individual gains an advantage, it is better able to differentiate odors, there will already be more loci in the odor recognition decision space. But at the same time, there is a problem with the decreased generalization ability of this system. The loci have become smaller, which means that more training examples are required to cover all spectra.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/431/1*8rxpPh79m0ebW-VyvPUZVg.png" /></figure><p>Evolution chooses the optimal value of the parameters of this system: a sufficient number of reference odor vectors greater than the number of required classes and yet small enough for a good level of generalization.</p><p>Evolution from species to species does not reinvent the mechanics, those solutions that worked evolve and are supplemented, so everything that we learned about the fly nervous system can be transferred to more evolved mammalian nervous systems as well. The obvious disadvantage of “fly solutions” is that the reference vectors of odor patterns are placed almost randomly during the formation of the organism, and their location most likely does not change during the course of life. The next improvement could be the ability to move the reference vectors of the patterns during learning, this would optimize recognition and make it possible to learn to distinguish between any two odors within the same individual.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/128/1*C-u8UaTM7f7fO_vakCv4Jw.gif" /></figure><p>Essentially, statistical learning is a process of preparing and tuning reference vectors of images as anchors. In machine learning, there is already a class of neural networks that are very similar in their operation to biological analyzers — these are <a href="https://en.wikipedia.org/wiki/Self-organizing_map">Kohonen neural networks</a>, the author of which is the Finnish scientist <a href="https://en.wikipedia.org/wiki/Teuvo_Kohonen">Teuvo Kalevi Kohonen</a>.</p><h3>“Biological” neural networks</h3><p>Such a method as modeling helps to understand many processes and phenomena in a very deep way. Let us also understand the living-learning process by the example of a neural network model configured to recognize handwritten digits from the standard MNIST set (as well as the fashion-MNIST pentagram of clothing items).</p><p>Our network has an input in the form of an array of points 28*28, each point a number from 0 to 1 in a gradation of 256. The network consists of two layers: the first layer is conventionally called the Kohonen layer, the second layer is the Hebb layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*XZ0xpJrgZT_OyAX6gBArLA.png" /></figure><p>The Kohonen layer is represented by an array of 32*32 neurons, each such neuron is connected to all inputs of the network. Each such connection is defined by weight, a real number from 0 to 1. All weights will be stored in one array of size 896*896 (32*28=896), I use for this purpose images, for simplicity and clarity, so let’s call this array a map of synaptic connections or a map of weights. The map of weights will initially be filled with random values. When an image is fed to the input of the network, the activity level of each neuron of the Kohonen layer is determined by the proximity of the input vector and the vector stored in the weights of the neuron.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/163/1*D_uDG_a77sLPdROgRQwr1Q.png" /></figure><p>where Ai is the final activity of the i-th neuron; Wij is the j-th weight of the i-th neuron; Ij is the j-th input element; n is the vector length.</p><p>Thus, it turns out that the closer the input vector and the vector stored in weights are, the closer the neuron activity is to 1. The Kohonen layer works on the principle “winner takes all”, so we select the neuron with the maximal value of activity and train it by bringing the vector of its weights closer to the input vector by a certain value. For faster learning you can choose the maximum coefficient of change (Rate) of vector during training, thereby the winning neuron will immediately assign itself the value of the input vector.</p><p>And here we see one of the problems, which is associated with the “winner takes all” learning. If the initial random vectors turn out to be far from the example vectors, the training may produce a neuron leader, which will always take the win.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/267/1*Gizu064arU7EbZ4D2HBqsQ.gif" /></figure><p>There are several solutions to this problem. First, it is possible to fill initially the weight map not with random values, but with vectors from the training sample. This method seems cheater to me because the newborn system does not have such a possibility as to obtain images of the external environment before interacting with it. Secondly, the way used in self-organizing Kohonen maps is to train not one winning neuron, but its neighboring neurons in a smaller decreasing degree with increasing distance from the winner. Gradually you can decrease the training radius. This method allows you to create some kind of image map, where you can see a smooth transition from one image to another.</p><p>Example of a Kohonen map on MNIST examples</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fd550hWUhppk%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dd550hWUhppk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fd550hWUhppk%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/af81f6272062946b3f0a5f59567096f7/href">https://medium.com/media/af81f6272062946b3f0a5f59567096f7/href</a></iframe><p>But our task is not to create such maps with smooth transitions between vectors, we aim at the specialization of neurons, at obtaining the effect of “granny neurons”, i.e. it is desirable to get as many diverse specific vectors as possible. We will use the principle of neighborhood learning only at the beginning of training, as it contributes to faster learning.</p><p>In my article “<a href="https://blog.goodaudience.com/comparison-of-the-brain-with-a-computer-neural-network-3ecd3c1fef57">Comparison of the brain with a computer neural network</a>” I expressed the idea that a neuron strives for a certain balance in the number of its activations, lack of activations is not desirable, neither is excessive activity. Let us apply this theory to practice. Each neuron of the Kohonen layer will have some parameter of involvement in the learning process. Each time a neuron wins, this parameter usually increases by a rather high level, and when it loses, this parameter will insignificantly decrease. And this parameter will affect the probability of the neuron’s victory, the higher the involvement in learning, the lower the chance to win. That is, the victory of a neuron is determined not only by the proximity of the vector on the neuron weights and the input vector but also by how long ago or how often the neuron won. Thus, neurons are involved in training more evenly throughout the layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/655/1*TlPV7i3HPwWX5LxSIRsorw.gif" /></figure><p>After presenting the network with 2,000 random examples, we will fill all weights of all neurons of the layer with different examples, here examples of different digits with different outlines are highlighted. To see the result of the layer in runtime mode, we do the following: first, we determine the activity level of each neuron in the Kohonen layer using the formula described above, and second, we determine the maximum and minimum activity values in the layer. Then we calculate for each neuron the ratio of maximal and minimal values according to the formula (fact — min)/(max — min). Thus, the minimum value is equated to 0, and the maximum value to 1. Already in this way we can clearly see which neurons are maximally close to the input vector, and which ones are maximally distant. Since the obtained values are in the range from 0 to 1, we can make this picture even sharper, more contrasting just by taking this value to a certain power (I have the power of 15 in my examples). The maxima will then remain unity, and values much smaller than unity will become almost zero. This operation is called escalation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/310/1*GP1Cw9kJremGPbK-nDCLXw.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/631/1*jbZeIP6nM0t2v8JqLkbWEA.gif" /></figure><p>It turns out that if we feed a vector to the input of our Kohonen layer, we will get the activity of the specialized neuron closest to the input vector at the output. It remains to match this activity with the required corresponding class.</p><p>This is done by the second layer of the network — Hebb’s layer. The Hebb layer has 10 neurons by the number of classes, the weight map of this layer has the size 32*320, i.e. all neurons of the Kohonen layer are connected with all neurons of the Hebb layer. Learning of this layer is rather trivial: when the activity of neurons from Hebb’s layer and Kohonen’s layer is combined, the connection weight becomes stronger and weakens when there is no combination. When performed, the weights are simply summed up, and as in the Kohonen layer, only one neuron from the layer wins. The winning neuron is the response of the network.</p><p>These two layers can be trained together or separately. For example, we can present 2,000 random examples by training the Kohonen layer and simultaneously pass information about the class of the example to the Hebb layer by training it. For my implementation in this mode, 2,000 examples are sufficient for a result in the neighborhood of ≈82% quality (the test runs on 10,000 test examples).</p><p>Another example of learning, which is closer to biological learning, is when the Kohonen layer learns from random unlabeled examples. Thus forming basic images. Then you can present the network with several marked examples from 10 to 1000 and train only the Hebb layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*acPK0c-QQrTg1eg9Fay_rQ.jpeg" /></figure><p>The network has been trained on 2,000 unlabeled examples and it is sufficient to train it on 300 labeled examples for quality above 80%.</p><p>The marked data may be small relative to the obtained reference images and not all can be involved, in which case a good option would be to reduce the degree of escalation.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FD4SdCz4x_fE%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DD4SdCz4x_fE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FD4SdCz4x_fE%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d50ed814772f38d1fe7ecbbe8d1ebc15/href">https://medium.com/media/d50ed814772f38d1fe7ecbbe8d1ebc15/href</a></iframe><p>Examples of training on the MNIST dataset are given. It is possible to download and run the applications at the links below, as well as to see and, if desired, use the source code of the network implementation in the Unity3D environment.</p><p>Source: <a href="https://github.com/BelkinAndrey/BrainSquared">github.com/BelkinAndrey/BrainSquared</a></p><p>Download build for Windows: <a href="https://drive.google.com/file/d/1rJk_YVvu-lkRXXcZ1Wb-ISw3S8kYCjEG/view?usp=sharing">MNIST</a>; <a href="https://drive.google.com/file/d/1HAFFjH_SYrn6aAtmKY4wYtM0173M7FCt/view?usp=sharing">Fashion-MNIST</a></p><h3>In comparison</h3><p>The described network has a very significant disadvantage — it’s low quality. The learning quality of this network will always have a certain ceiling, which depends on the total number of reference images stored in the Kohonen layer, the more neurons of this layer, the possible level of quality is higher. But with an increasing number of neurons more examples are required for their training.</p><p>Backprop training and biologic network training:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*EhIyYvBMXRBjIxXYCHd0pQ.gif" /><figcaption>Backprop training</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*yXw2O-KEH0yfz5FKZJ2HSA.gif" /><figcaption>biologic network training</figcaption></figure><p>Learning the network through backpropagation of error allows you to draw the boundary between classes very accurately, the more examples, the more accurate the boundary. The boundary between classes in the training of the Kohonen network passes only along the loci boundaries, which creates inaccuracies. Therefore, even if it is possible to move reference vectors, the limitation in their number may not allow achieving very high quality.</p><p>BUT! The described type of network has a number of advantages, which potentially makes it possible to create artificial intelligence based on them in the future. First, the ability to abandon backpropagation. Backpropagation of an error is something alien to biological nervous systems. Second, it is based on self-learning, an analog of statistical learning; the trend towards embedding in self-learning is very relevant in machine learning. Third, the ease of computation, which is very relevant to the rising prices of computational resources. Related to this is another fact: The brain is not a neural network! The brain is a multitude of neural networks!!! With such a concept, 80% of the network quality will be compensated by their quantity. Imagine that a number will be “looked at” not by one network, but by several at different angles and scales, and the final value is determined collegially. And the ease of computation for such configurations will play a role.</p><p>For evolution, speed of learning played a major role, not accuracy. Humans cannot use a picture of a mole to diagnose a tumor with a 99.8% probability. Accuracy is the domain of machines, not of Man. The described network is closer to biological networks than neural networks trained on the basis of backpropagation.</p><h3>In conclusion</h3><p>Kohonen networks have been known for more than 40 years, but they have not developed as rapidly as the networks trained on the basis of backward error propagation algorithms. But there is more biology in the way these networks are trained than in the currently popular machine learning methods. At the moment I am working on the development of such networks, primarily on the task of creating “deep”, multilayer networks. Such models need to learn from the bottom up, rather than from the top down, as is the case with backward error propagation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*qnQyIHdGi08_adAtdssBlw.gif" /></figure><p>After all, our brain and its visual analyzer in particular is capable of passively observing learning and creating high-level, so-called embeddings, features, and these embeddings will be associated with the activity of specific neurons. And the brain doesn’t use a backpropagation and needs fewer examples.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=597e1ba2fc59" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/living-creature-learning-and-the-computer-neural-network-597e1ba2fc59">Living creature learning and the computer neural network</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A super-efficient neural network or hypernet]]></title>
            <link>https://medium.datadriveninvestor.com/a-super-efficient-neural-network-or-hypernet-67d6be300a7?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/67d6be300a7</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[brain]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Thu, 17 Dec 2020 02:34:37 GMT</pubDate>
            <atom:updated>2020-12-26T13:52:51.543Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fXyL403_aSTlG0NQUVUPvw.jpeg" /></figure><p>What does it take to create artificial intelligence? Which way to go to this goal? Many will confidently answer that the scientific community is already moving at full speed along this path. That with the next neural network models with even more parameters, with even more powerful neuromorphic computers, and even larger datasets for all occasions, we will break into the era of thinking machines. In my opinion, it looks like a race for a carrot and this state of affairs will become more obvious to everyone when the number of tunable parameters in the models becomes more than connections in the human brain, but the expected effect will not be achieved.<br>In this article, I will outline the way out of the vicious race for carrots, and talk about my path and developments in my research.</p><p>The first thing that needs to be addressed on the way to artificial intelligence is the problem of low training efficiency of neural networks. It is foolish to build a grand high-rise building without a foundation, or on a very bad foundation. Many researchers continue to develop their concepts and models of intelligence on this foundation, instead of focusing on the obvious problems: biological neural networks learn many times more efficiently than artificial neural networks.</p><p>Yes, biological neural networks, one might say, are in a constant continuous flow of information, continuously receiving signals from their receptors. In a day, an infant can receive a gigantic amount of information from his or her organs of sight and hearing, equivalent to video and audio data for many, many terabytes. But if we analyze this information in detail, then there is not so much information that is really useful for training, so to speak, “marked” information. We can say that a person learns on a very poor quality dataset, most of the examples are irrelevant and not marked up, some of the examples are very noisy, but at the same time, he learns quite successfully and quickly.</p><p>Apparently, computer neural networks do not correspond to biological ones or we train them in a completely different way from the way it happens in the biological nervous system.</p><p><em>It is necessary to concentrate on developing super-efficient neural networks, rather than increasing the number of learning parameters.</em></p><p>Moreover, the main focus should be on the speed of network learning, while quality or accuracy is now secondary.<br>Only a super-efficient neural network (SNN) can be the foundation and foundation for creating artificial intelligence.<br>Having SNN technology, you can move in different ways, like creating cognitive architectures based on graphs, hypergraphs, etc. plus SNN, and moving in a biological way repeating the architecture of the biological brain. I am a supporter of the biological approach, although I do not consider this path to be the only one in achieving the goal.</p><h4>Roadmap towards AI</h4><ol><li>First stage. Creation of a model of sensory analyzers, primarily visual and auditory. In general, sensory analyzers of the biological nervous system should become the prototype of super-efficient neural networks (SNN).</li><li>Second stage. Gaining knowledge about the mechanisms of instant, temporary, and long-term memory in the concept of SNN. Biological sensory analyzers easily work with context, only understanding the concept of memory in biological neural networks will allow us to adopt the ability to work with context, which will greatly expand the capabilities of the SNN.</li><li>Stage three. Creation of models of motor networks. Despite the homogeneity of the <a href="https://medium.com/datadriveninvestor/modeling-of-the-cerebellum-c841023a4dcf">cerebral cortex</a>, motor networks are in some ways more complex than sensory analyzers, although there are some deep similarities. Therefore, this stage will not belong. The motor networks will include a model of the cerebellum, as a structure that memorizes and reproduces short time intervals for precise movements. And also a kind of reward mechanism to ensure the learning of correct movements, similar to the dopamine reward mechanisms in the biological brain.</li><li>Fourth stage. One of the important forms of motor actions is speech. Perhaps, for this reason, this and the previous stages will be implemented simultaneously. At this stage, the speech will not be fully conscious. As a person, stages of speech development must be passed, such as humming, the pronunciation of individual sounds, then words. The maximum level of speech proficiency at this stage is naming the presented objects, describing scenes, answering simple questions, reading, or declaring the text. At this stage, the foundation for the network of the speech circle will be laid (I wrote about the speech circle in detail in the article “<a href="https://medium.com/datadriveninvestor/what-is-the-mind-b59d639b6c06">What is the mind</a>”).</li><li>Fifth stage. Creation of a model of <a href="https://medium.com/datadriveninvestor/simulation-of-emotions-2e01b2dc2b2e">basic emotions and needs</a>. This stage will not be a clear-cut stage; many of its concepts will manifest themselves at earlier stages. The main fundamental emotion that needs to be modeled is the emotion of novelty. It is enough to create an evaluative mechanism of information novelty at various levels of abstraction and to perform modulation of network parameters on its basis. Curiosity is essential for effective learning. You also need to provide the models with the senses they need to control their needs. For example, a low battery in a robot could create a certain feeling of hunger for the robot, and the robot would seek to satisfy it, that is, it would look for a charger or power source. Or creating some kind of feeling of pain or discomfort so that the robot avoids traumatic situations. Another useful emotion that can be easily modeled is the attachment of various kinds. Affection is based on an imprinting mechanism. This will be useful for all sorts of helper robots or pet robots that are required to follow a master&#39;s and have devotion. A person will be more willing to interact with intelligence similar to him, therefore, a wider range of emotional mechanisms may appear. For example, anxiety and fear, sympathy, joy, and disappointment. As for the introduction of Isaac Asimov’s laws of robotics to the level of basic emotions, they have a high abstract character, which greatly complicates their modeling. First, the model must be trained so that it is possible to identify the representation of the images “Human” and “harm”, similar to how one can find “grandmother’s neurons” in a person. Then you can associate the emotional mechanism with the desire to avoid combining these images. It is difficult to say how effective such a method will be.</li><li>Sixth stage. We can say that this stage is as difficult as the first, if not more difficult. Conditionally, it can be called the stage of the prefrontal cortex. Higher cognitive activity is assigned to the prefrontal cortex; we can say that judgment and long-term goal-setting are located. The prefrontal cortex controls the work of the emotional centers and centers of need, it regulates how much they influence behavior. Working at this stage requires in-depth knowledge of human behavior. Only at this stage can one reason and talk about consciousness and conscious behavior. To this stage, I include not only the creation of the prefrontal cortex but also several important biological neural networks and mechanisms. For example, the “interpreter” network, which seeks to justify any of our actions. The presence of such networks is clearly evident in experiments involving people with split-brain hemispheres.<br>Or the network responsible for the illusion of qualia, creating a sense of centralized perception of the world. Also, the mechanism of perceiving the boundaries of one’s body and actions, a violation of this mechanism in people can lead to an obsessive desire to get rid of a part of their body or the inability to distinguish their actions from those of another person.<br>We know that disturbances in various mental mechanisms can lead to various diseases on the autism spectrum or various forms of schizophrenia. On the way to artificial intelligence, it is destined to encounter examples of these deviations in the artificial psyche, but it also has a useful grain. Work on artificial intelligence will allow us to better understand the nature of human mental illness.</li></ol><h4>Hypercolumn network, or hypernet</h4><p>We are now at the first stage of this roadmap, at the stage of creating super-efficient neural networks. For me, this is not quite the first stage, first I had to understand the principles of the peripheral nervous system and how neural networks work in <a href="https://medium.com/@it.belkin/opentadpole-the-first-cybernetic-animal-1dcba8a01e9d">simple nervous systems</a> and, in general, create more than one model of the operation of neurons. Now I am developing a model for the main structural element, which is not a neuron, but a hypercolumn or cortical module. Hence the working name hypercolumn network or <em>hypernet</em> for short.</p><p>Next, I will tell you in more detail what principles it is based on and how I implement it.</p><h4>Second signaling system — nodes</h4><p>Let’s see how information from receptors can be represented in the cortex, for this, we need to understand how two simple mechanisms of irradiation and induction, known for over a century, proceed.<br>Irradiation is the tendency of nerve excitation to spread to neighboring neurons or neighboring nerve tissue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*8TMkYNvW01mTFpYeYuz6ug.jpeg" /></figure><p>Usually, the processes of cells of one column can spread to adjacent columns and exert a stimulating effect in adjacent columns. The nature of such an incentive effect on adjacent columns can be represented as a certain circular halo with a maximum effect in the center and with a gradual weakening with distance from the center.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*2jWFgTqdfGbQopEboiJuTQ.png" /><figcaption>Irradiation map</figcaption></figure><p>Induction is the appearance of areas of inhibition on the cortex around the site of excitation. This may be because outgrowths of neurons of one column at a certain distance of the column can form connections with inhibitory cells of other columns. The nature of the inhibitory effect on adjacent columns can be represented as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/1*LIK6d8fsVVlFfte2pVA2rA.png" /><figcaption>Induction map</figcaption></figure><p>These processes work in the cerebral cortex at the same time. Let’s simulate how the excitation arising in the cortex under the influence of these two mechanisms will transform. For this, a pixel shader was written — a program executed by the GPU for each pixel in an image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/682/1*MLiZb8X2kMdD6fACI1nZQQ.jpeg" /><figcaption>Shader in Unity</figcaption></figure><p>In the model, each pixel of the image will be a separate cortical module; the brighter the pixel, the stronger the overall excitation in the column.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*NVfuTp37POmvP3z7NyaeBg.gif" /><figcaption>An example of the shader, the image size is 28x28 pixels, the radius of the column is 5 pixels, the shader is applied several times</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/208/1*iSIMNteo1B5rtpse8fPupQ.gif" /><figcaption>Column radius of 7 pixels</figcaption></figure><p>As you can see, under the action of the mechanisms of irradiation and induction, any activity is broken down into small foci of excitation. The size of these foci depends on the size of the total impact on neighboring hypercolumns, as well as on the ratio of the forces of action of incentive and inhibitory maps.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*FjcAn5dQZXjmcNzuSQ-tUA.jpeg" /></figure><p>These foci of excitation are encapsulated, as it were, surrounding themselves with an area of ​​inhibitory action. Also, their central part, with several applications of shader calculations to the image, due to irradiation, becomes more and more active, thereby these foci become pronounced and stable.</p><p>These centers of excitation will hereinafter be called <em>nodes</em>. Nodes are some new emergent property of the cortex; it manifests itself as a result of the mutual work of two mechanisms of irradiation and induction. The nodes are not tied to the size and position of either neurons or hypercolumns, they work as if on these structures. The combination and location of nodes is the representation of information in the cerebral cortex.</p><p>Nodes are the second signaling system, symbolic representations of images in the brain.</p><p>It is amazing how simple and elegant solutions are used by the cortex and the brain as a whole.<br>To understand how nodes can store images or parts of images, it is necessary to study another mechanism at work in the cortex.</p><h4>Reverberation of excitation</h4><p>The reverberation of excitation is a prolonged circulation of excitation impulses in networks of neurons or between different structures of the central nervous system. For us, corticothalamic reverberations and reverberations that occur between different areas of the cortex are of great importance. It is these reverberations that create various patterns and rhythms in the electroencephalography of the brain. For such reverberations, descending or feedbacks is required, which are usually several times more than ascending direct connections.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/831/1*Qe0Y1uZfezBFcxwOKP54cQ.jpeg" /></figure><p>We need four maps to simulate the reverb process. Here I refer to the images as maps, to which we will apply some pixel shaders, if necessary for our model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/639/1*HLpgJtHiIZp2UR0zyb3Lkw.jpeg" /></figure><p>A map of the thalamic nucleus onto which we will project some images, in this case, examples of handwritten numbers from the standard MNIST set. The images in the MNIST dataset are 28x28 pixels, so we’ll set this map to 28x28 pixels as well.<br>The cortex map will be 21x21 pixels in this version.<br>The map of ascending synaptic connections will store the values ​​of all weights of the synapses connecting the thalamic nucleus map and the cortex map. Conventionally, each pixel-hypercolumn of the cortex will have synapses with 64 (8x8 nucleus) pixels of the thalamic nucleus map. If you move the 8x8 pixel kernel along with the thalamic nucleus map in one-pixel increments, then you can make a map of the ascending synaptic connections. Accordingly, this gives the size of the cortex map 28–8 + 1 = 21, in turn, the size of the map of ascending synaptic contacts will be 168x168 pixels (21 x 8 = 168).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JbvezTlwc3TaUkfeLPtDvA.jpeg" /></figure><p>The synapse weight will be determined by a number from 0 to 1 — this is convenient for graphical display on a map and allows you to store and process these parameters in the form of images, which is also convenient in working with shaders.<br>Another important point, in the models that I will describe, there are always two maps of synaptic connections: one map of incentive synapses, the second map of inhibitory synapses. These are always two cards of the same size.<br>It is known that the type of synaptic contact in the biological nervous system does not change during life, that is, an incentive synapse cannot become inhibitory, or vice versa. In computer neural networks, the sign of the synapse weight can easily change during the learning process, which is not correct for a model claiming to be biological.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n8Ck8PEvRkzeENrhyWL6Jg.jpeg" /></figure><p>In the columns of the cortex, in addition to the incentive pyramidal neurons, there are many inhibitory neurons, and any axon from the incentive cell directed to the column can create synaptic contact with both the incentive and inhibitory cells. Similarly, we have two synaptic contact cards, one will act with a plus sign, the other with a minus sign. In both of these cards, the weight will be determined by a number from 0 to 1.<br>We fill the value of the weights of the ascending synaptic contacts with random numbers from 0 to 1.<br>The map of descending synaptic connections will be significantly larger. As already noted, there are several times more descending connections in the biological brain. In our case, we choose a map size that will demonstrate the concept of storing images in nodes. Each pixel-hypercolumn of the cortex map will be associated with each pixel of the thalamic nucleus map, that is, with every 784 pixels (28 x 28 = 784), then the map size will be 588x588 pixels (21 x 28 = 588).</p><p>Initially, the value of the weights in the map of descending synaptic connections will all be equal to 0, but we will change these weights during training.<br>The teaching method that we in this case apply to the nervous system has been experimentally confirmed, this is the Hebb rule. Its simplest interpretation: we work together, it means we strengthen the connection; as well as the downside: if they didn’t work together, the connection is weakening.<br>First, let’s give our model an example by projecting it onto a map of the thalamic nucleus. Then, through the upward connections, we have a summing effect on the map of the cortex. For each pixel of the cortex map, all the corresponding weights from the map of ascending synaptic connections are added, multiplied by the value of the activity level of the corresponding pixels of the thalamic nucleus map. The value of the weights from the inhibitory map of the ascending synaptic connections will be added with a minus sign. The resulting amount is multiplied by the selected coefficient, called the sensitivity coefficient. In this case, it is 0.5.<br>If the received sum is less than or equal to zero, then the pixel-hypercolumn is not activated, its activity is equal to zero. If the resulting sum is greater than one, then the pixel value of the bark map will be equal to one, like the maximum degree of activation of the hypercolumn. If the sum has a value in the range from 0 to 1, then the pixel-hypercolumn of the bark map will take that value.</p><p>The sensitivity coefficient is selected so that as a result of the sum, a very large number of hypercolumns of the cortex map is not activated since this leads to mutual suppression of activity when using induction. Also, the sensitivity factor should not be too low, because this can lead to the very low activity of hypercolumns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*55ZWvagOge12NRIyxx0qoQ.jpeg" /></figure><p>Next, we apply an irradiation and induction shader to the cortex map. We can apply this shader several times. As a result, we get several nodes on the cortex map.<br>Then we turn to the map of descending synaptic connections and apply the Hebb rule for learning to it. And we use it in the following way: in the case of joint triggering, we significantly strengthen the synaptic connection. If the presynapse works, but the postsynapse does not, then we weaken the synaptic connection only by a small amount. In this case, with joint triggering, the weight of the synapse will increase by 0.8; in the absence of joint triggering, the weight will decrease only by 0.1.<br>Since it is the coincidences that matter more than their absence. Note that the presynapse is tuned, not the postsynapse, although now this is not critical, since the entire synapse is regulated by one weight, but this has an important ideological significance. The fact is that in classical neural networks the emphasis is on the tuning of postsynapses, and changes in presynapses are completely ignored, although in biological neural networks changes occur in both parts of the synapse, and in the presynapse, they are often more pronounced (<a href="https://blog.goodaudience.com/comparison-of-the-brain-with-a-computer-neural-network-3ecd3c1fef57">Comparison of the brain with a computer neural network</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/849/1*d9PBDgswKOLiynUtcMbX-A.jpeg" /></figure><p>For an inhibitory map of descending synaptic connections, it is logical to apply Hebb’s rule, on the contrary, the inhibitory connection will increase when the postsynapse is not activated when the presynapse is active.<br>This is how the maps of descending synaptic connections look like after training using one hundred examples (ten examples per figure).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ckB2iNjLjtDxef1PkRamRw.jpeg" /></figure><p>You can see that the maps of descending synaptic connections reflect images of whole numbers, and in those places that correspond to the nodes that are activated on these images. This is possible because the selected number of feedbacks for each microcolumn covers the entire map of the thalamic nucleus; if the feedbacks covered only a part of the map, then the hypercolumn could store only a part of the image. So in the nodes, some images or some component parts of images can be written.<br>When activated, the node “looks” at the activity that activates it and keeps it in itself. The nodes store, as it were, some reference examples of images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/457/1*oE4cQ6YfRsX7_S2yzklTZA.gif" /></figure><p>Now, after learning the map of descending synaptic connections, you can start the cyclic transmission of excitation — reverberation.<br>We can see examples of how the initial projection of an example onto a map of the thalamic nucleus will be converted in the process of reverberation into a certain reference version of the same figure, and with this version, the reverberations will be stable.<br>This process is very similar to the work of Hopfield networks, due to the presence of recurrent connections, this type of neural networks, in my opinion, is closest to the work of biological neural networks. In contrast to Hopfield networks, in our version, there is a certain separation into nodes, which gives some advantages. The number of inputs and outputs in Hopfield networks is the same, each neuron is connected by a recurrent connection with other neurons. It turns out that each image is, as it were, stored in the weights of each neuron, which greatly reduces the capacity of the network. In our case, separate groups of columns are selected to store individual images. Our version of the network is more interpretable, we can say exactly for which image this or that element is responsible. Here we can trace an explicit representation of information, similar to the presence of “grandmother’s neurons” in the cerebral cortex, we see the pixels-columns of the cortex map responsible for the image of a particular number.</p><p><a href="https://www.datadriveninvestor.com/2020/11/19/how-machine-learning-and-artificial-intelligence-changing-the-face-of-ecommerce/">How Machine Learning and Artificial Intelligence Changing the Face of eCommerce? | DataDrivenInvestor</a></p><h4>Modeling the effect of LSD on the cerebral cortex</h4><p>This small digression will demonstrate some similarities between our model and the work of the biological brain. The fact is that for us the effect of LSD on the bark is very easy to simulate. LSD by its action interferes with the work of the neurotransmitter GABA, which is the main inhibitory neurotransmitter of the cortex; it works in the induction mechanism. Therefore, in order to simulate this, it is enough to reduce the degree of induction in the settings of our pixel shader.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*NhXu87cqfu-TK_eTzuEitQ.gif" /></figure><p>As we can see at the same time, the usual borders, and the sizes of the nodes are blurred and expanded. If the effect of induction is further reduced, then the foci of activity could merge, forming some patterns. Similar phenomena can occur in the cerebral cortex under the influence of LSD, which is indirectly evidenced by MRI scans taken from people who took LSD, the activity of the cortex, in this case, is much larger and more extensive than during normal brain work. Considering that almost every column contains images or their parts, it becomes clear why there is confusion and mixing of images, the reasons for the impairment of perception are clear. Such disturbances in the functioning of the brain cannot be useful for him.</p><p><strong>Drugs are bad for brain health and health.</strong></p><h4>Learning ascending synaptic connections</h4><p>The tuning of the ascending synaptic connections is also of great importance since their tincture determines where the nodes will arise when using irradiation and induction.<br>To begin with, we will slightly modify and complement our model. First, we will increase the size of the cortex map; in the biological brain, the size or area of ​​the cortex is much larger than the size of the subcortical and thalamic nuclei with which it is associated. Secondly, we will add another layer of information processing, since the higher levels are of great importance in teaching the upward linkages of the levels below.<br>To increase the size of the cortex map, while maintaining the size of the thalamic nucleus map (28x28 pixels), we will have to increase the number of ascending synaptic connections. To do this, we will apply some kind of nerve model, which is a connection map. A nerve map is an image that uses two-color channels to encode the pixel coordinates of the signal source map.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hWpKQTcJ1-uFtX3u8Bbf9w.jpeg" /></figure><p>We can create a nerve map to begin with, in which the coordinates of the pixels of the source map are consistent with the nerve map itself. And if we apply such a nerve map to our thalamic nucleus map, we just get an enlarged image. But if mixing of pixels is introduced into this map of nerve connections, then using such a map will lead to the fact that each pixel of the source map is represented on the map of the nerve as a set of pixels and, depending on the degree and method of mixing, the location of these pixels will be different. With mixing, in which the pixels are placed together at a short distance, you can get a kind of random accumulation of representative pixels. This is very similar to how the fibers from the axons of neurons from the lateral geniculate body fold into nerves, in which the axons are repeatedly separated, and the ends of the axons are dispersed throughout the primary visual cortex while maintaining some topography of the retina and at the same time mixing.</p><p>In our version, on a 28x28 pixel map of the thalamic nucleus, we will apply a 512x512 nerve connection map. This nerve map will connect the thalamic nucleus map with a 64x64 pixel map of the cortex, that is, break the entire nerve connection map into 8x8 pixel squares, then each such square will include all connections by a pixel-hypercolumn (512/8 = 64).</p><p>Nerve connectivity maps do not change during the learning and operation of the network; it is set before starting to work with the network, similar to how nerves are formed in the brain before birth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*20hRXBUHqA1moHuzAjNGRQ.jpeg" /></figure><p>Accordingly, maps of ascending synaptic connections will have a size of 512x512 pixels.<br>The map of downstream synaptic connections will have a size of 1024x1024 because the number of recurrent connections must be greater for stable reverberations. Here, each pixel of the cortex map will correspond to 16x16 pixels of the thalamic nucleus map (64 x 16 = 1024).</p><p>The next level of information processing is represented by the output map, which has a size of 1x10, one pixel per digit since our network is aimed at recognizing handwritten digits from the standard MNIST set. The exit map is related to the cortex map in the same way as the thalamic nucleus map is connected to the cortex map. Accordingly, there are also maps of ascending synaptic connections and maps of descending synaptic connections, which have a dimension of 64x640 pixels. Thus, each pixel of the cortex map has synaptic contact with each pixel of the exit map, so that the pixels of the exit map has a maximum receptive field. Descending synaptic connections, as already noted, usually have a greater number than ascending ones, but in this case, they will be the same in number, because this will be the maximum size of backward synaptic connections. So every pixel of the exit map “sees” every pixel of the cortex map.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/1*T00DSH-JI5H1bxOHzNmBIQ.jpeg" /><figcaption>Screenshot of the program</figcaption></figure><p>Training or tuning the map of the ascending synaptic connections of the second level can be carried out according to the Hebb rule, which we have already applied. In this case, during training, we will consider active or equal to 1 pixel of the output map, which corresponds to a digit, an example of which is presented to the network. Thus, we seem to associate some activity on the cortex map with the required output and strengthen this connection. This is similar to the formation of a reflex arc of a conditioned reflex, as the representation of a digit is linked by connections with a combination of active pixels in the thalamic nucleus map. Thus, elements that are actually not connected by direct contacts due to recurrent connections can connect and strengthen their connections among themselves. And it can work through many levels.</p><p>But if you train or adjust the map of ascending synaptic connections of the first level according to the Hebb rule, this rule leads to the fact that weakly active pixel-hypercolumns, due to the strengthening of connections with them, will increase their activity to the maximum, and this will create a large number of active pixel-hypercolumns, which leads to mutual suppression when using induction. And this leads to a strong decrease in the quality of the network.<br>Therefore, we will use a different teaching method. It will consist of two stages. The first stage is the initialization of images. At this stage, we will form a certain basic idea of ​​each recognizable digit pattern, using a small number of examples. For this, 10 to 100 examples are sufficient, or 1 to 10 examples per digit. At this stage, all synaptic connection maps will be configured except for the first level ascending synaptic connection maps, but they must be filled with random values.</p><p>First, we project the example onto the thalamic nucleus map, then summarize using the random weights of the synaptic map. Before applying the irradiation and induction shader, we can influence the cortex map so that the resulting nodes do not match for different numbers. To do this, let us turn to the map of downward connections of the second level, which stores “prints” of patterns of nodes for each digit formed on the basis of previous experience. Accordingly, if there was no such experience, then no influence will be exerted on the map of the cortex. For example, if the example is the number “5”, then to the map of the cortex we will add the values ​​of the incentive map of the descending synaptic connections of the sector corresponding to the image of the number “5”. This will increase the chances of knots forming in the same place where these knots were formed in the last act of learning the number “5”. We will also subtract the values ​​of the remaining sectors not associated with the number “5” from the bark map. This helps to ensure that nodes will not form in those places that correspond to other numbers.<br>After such a correction of the values ​​of the cortex map, we apply the pixel shader responsible for the irradiation and induction.</p><p>Then we set up the maps of the ascending synaptic connections of the second level according to the Hebb rule, taking into account that the pixel corresponding to the presented example will be active on the output map.</p><p>Also, according to the Hebb rule, we configure the maps of descending synaptic connections of the second and first levels.</p><p>This is what synaptic connections look like after such training using 100 examples:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_AD8S9KHiu1khfavMbGybA.jpeg" /><figcaption>a and b maps of the ascending incentive and inhibitory synaptic connections of the second level; c and d maps of descending incentive and inhibitory synaptic connections of the second level; f and g maps of the descending incentive and inhibitory synaptic connections of the first level</figcaption></figure><p>I pay special attention to the map of descending synaptic connections of the second level, on it, as it were, the positions of all nodes on the map of the cortex for each digit are captured. Moreover, the nodes for individual digits rarely coincide, that is, each digit has its own unique pattern of nodes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*6FCZVHF-aUtrcvpmiFVjgw.jpeg" /></figure><p>We move on to the next stage of training, in which we need to mainly adjust the maps of the ascending synaptic connections of the first level and, preferably, in such a way that the example presented to the unfamiliar network forms on the map of the cortex a pattern of nodes corresponding to the figure of the example. To do this, after summation, we will receive the difference between the values ​​of the cortex map and the corresponding example sector of the map of descending synaptic connections. Due to this difference, it is possible to adjust the weights in the map of ascending synaptic connections so that the presented example leads to the activity of the nodes corresponding to its image.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FXUGH6JZY-10%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DXUGH6JZY-10&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FXUGH6JZY-10%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/cbc9914354db755acbcb1b6bb8ecd6e0/href">https://medium.com/media/cbc9914354db755acbcb1b6bb8ecd6e0/href</a></iframe><p>The video shows the network training on 100 random examples and a test on 10,000 test cases. Depending on the selected training examples, perhaps from randomly selected weights for the map of upstream synaptic connections, the network in such a test gives a result in the range from 40 to 60%. This is a very modest result, but this task requires a more complex network architecture. Numbers and signs are recognized by our brains at a fairly high abstract level. Accordingly, the model requires more levels. And also in the biological visual analyzer, some additional mechanism is used, a special filter from networks of neurons that react to stimuli of a certain orientation. This filter simplifies the processing of visual information. Adding such a filter should also improve the quality of our network. All these modifications require further research, on which work will continue. The purpose of the article and this example is to tell about some of the theoretical aspects of the model and the approach in general.</p><p>Usually, the artist does not draw pictures like a plotter, starting from the top corner, drawing all the details of the underlying work at once. For a start, the artist puts on the canvas a sketch of thin, barely noticeable lines indicating the general plan of the artistic concept. Gradually, the picture becomes overgrown with details, more and more small, at the end only the most accurate and barely noticeable brush strokes are applied. Similarly, you need to approach the issue of brain modeling. It is hopeless to try to model in detail by sequentially creating cortical modules with multiple synaptic connections with a wide variety of types of neurons and neurotransmitters. The best solution is to create initially simple simulations of individual mechanisms by gradually adding the necessary details to the models.</p><p>An attempt to simulate the interaction of two simple mechanisms in the cerebral cortex led me to the idea of ​​a hypernet described above. The development of the idea of ​​super-efficient neural networks will fundamentally change the state of affairs in the field of artificial intelligence. Such networks do not require a large amount of training data, which means that many tasks that are difficult to implement in machine learning due to the complexity of collecting training data will be solved, and the speed of development in applied problems will be significantly accelerated. Computing resources for training will not be required so high, which means that there is no need to build powerful computers with a specialized architecture, which is very expensive and laborious. Developers will become less dependent on technology, data, and services. Everyone will be able to create cool intelligent software without significant costs. Which will create an avalanche of interesting new technical solutions in the field of AI.</p><p><strong>Gain Access to Expert View — </strong><a href="https://datadriveninvestor.com/ddi-intel"><strong>Subscribe to DDI Intel</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67d6be300a7" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/a-super-efficient-neural-network-or-hypernet-67d6be300a7">A super-efficient neural network or hypernet</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Comparison of the brain with a computer neural network]]></title>
            <link>https://medium.com/good-audience/comparison-of-the-brain-with-a-computer-neural-network-3ecd3c1fef57?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ecd3c1fef57</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Sat, 25 Jul 2020 12:35:13 GMT</pubDate>
            <atom:updated>2020-07-25T12:35:13.660Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YVyAsmj-Mu791D8bX2Yoeg.jpeg" /></figure><p>One can come across many criticisms that the biological brain or biological neural networks work completely different from the current popular computer neural networks. Various experts resort to such remarks, both from biologists, neurophysiologists, and specialists in computer science and machine learning, but there are very few concrete comments and suggestions. In this article, we will try to analyze this problem and identify particular differences between the operation of a biological and computer neural network and suggest ways to improve computer neural networks that will bring their work closer to their biological counterpart.</p><h3>The knowledge boundary</h3><p>First, I want to explain why, in my opinion, everything is so bad in the matter of creating a strong artificial intelligence, despite the tremendous advances in computer science and knowledge about the biological brain. First of all, this is due to the large ideological gap between these two pillars of science. Computer science requires a certain schematic simplicity, rigor, and conciseness in the description of systems, a certain systematic approach, in discarding unnecessary information and clear structuring sufficient for design in the program code. In biology, the details of the description of the observed systems dominate; nothing can be discarded or ignored from observations. All observable facts should be included in the systems described. Therefore, it is difficult for biologists to apply a systems approach to their vast knowledge to create brain algorithms. To design an airplane you don’t need to repeat the image of a bird.</p><p>On the other hand, it is easy to understand scientists and engineers who, when immersed in the study of computer neural networks, from the description of the principles of brain operation, are content with a short paragraph of text about a neuron, without applying any critical assessment to this knowledge. A neuron with the help of synapses on the dendrites “listens” to other neurons and transfers the result of summation calculations further along the axon. Even neuroscientists use the formal McCulloch-Pitts neuron to describe the principles of a biological neuron, but they do it for another reason because there are no worthy alternatives, there is no clear description in the biology of what a neuron does, what logic it performs, despite extensive knowledge about him.</p><p>If someone tries to reengineer the work of the brain, they will come across a whole layer of accumulated conflicting knowledge. It is possible to work with such a volume of information only through the prism of some general theory of the work of the brain, which does not yet exist.</p><p>Humanity possesses technologies of colossal computing power and a gigantic amount of knowledge about the brain, but cannot obtain a synthesis of these things. Let’s try to solve this problem and erase this knowledge boundary.</p><h3><strong>Brain it should be simple</strong></h3><p>The first very important principle that should be followed is the idea that the brain should work according to some very simple rules, i.e. all cognitive processes, no matter how complex they may seem, are based on simple basic principles. This is different from what we are used to hearing about the brain. The long absence of a general theory of how the brain works have generated a lot of speculation that the brain is some incomprehensibly complex object, or that the nature of its work goes far beyond the scientific methods of study that are applied to it. For example, the brain is compared to a quantum computer, or the properties of complex computers are undeservedly attributed to individual neurons, which, together with their number in the nervous system, makes the requirements for computing power for brain modeling unattainable.</p><p>So what is the evidence for the simplicity of the brain? Here I will give a completely paradoxical example. If we take a grape snail and bring electrodes to one neuron of its large ganglion, according to all the requirements that apply to such experiments, then we can get a graph of the activity of an individual neuron, and try to analyze it, then we get a very complex nature of its activity. Even if we take into account the nature of the invasiveness of our experiment, then that our electrodes cause serious damage to the snail and limit its vital activity, the nature of the neuron’s activity still looks very complex. We will see both spontaneous activity and changes in the number and frequency of spikes over time. Many scientists have been struggling to explain this complex behavior of a neuron for a long time, looking for some pattern in this.</p><p>These facts make the neuron a kind of complex calculator, working according to a complex algorithm. Considering that there are about 20 thousand such neurons in the snail’s nervous system, we can say that the computing power of the nervous system of an ordinary snail is comparable to that of a mainframe. I think this should give you respect for these animals. But let’s see how complex the behavior of snails is. A snail is a kind of biological automaton, but it has a certain degree of variation in behavior, but it is very small. This is a set of unconditioned reflexes, often very simple, which can be explained by the already existing knowledge about neurons, synapses, and reflex acts, and there will be no place for complex calculations.</p><p>In support of the above, I want to make a reference to my last article, which describes a model of a frog tadpole, in which, thanks to the nervous system, the rather complex behavior of an animal can be obtained from several dozen neurons. Moreover, from very simple neurons, the model of which is based on facts known in science.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*wt-KJZPUbfRCJJ1XgXNBAA.gif" /><figcaption><a href="https://medium.com/@it.belkin/opentadpole-the-first-cybernetic-animal-1dcba8a01e9d">OpenTadpole</a></figcaption></figure><p>So where does this complex neuron behavior come from, and why are there so many? Here, in fact, one follows from the other. In nature, there is a paradoxical phenomenon that can be called the paradox of neuron efficiency. It lies in the fact that with the increase and complication of the nervous system, the efficiency or role of an individual neuron in this system decreases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DaEWHH02N0O7pdP7JWWE8w.jpeg" /></figure><p>If we analyze the nervous system of the c.elegans worm, whose connectome consists of 301 neurons, we will see that not only individual neurons are important in the proper functioning of its nervous system, but even individual synapses are important. That is, we can assign 100% efficiency to an individual neuron in the worm. If we consider the human nervous system from this point of view, it is difficult to assign to neurons a significant value of efficiency that can be taken out of the skull with a crowbar, while preserving the vital activity of a person and even his social integration, well, almost preserving. *</p><p><em>*</em><a href="https://en.wikipedia.org/wiki/Phineas_Gage"><em>a reference to the very famous case of Phineas Gage</em></a></p><p>Regularly, you can find articles that describe cases where people who are living full lives and are socially adaptive suddenly find that their <a href="https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(07)61127-1/fulltext">brain is devoid of any areas or lobes</a>.</p><p>If we observe the activity of a healthy brain, then we will not see any dormant neurons, each neuron will be activated, to varying degrees, of course, each will be assigned its own role. How the brain does this, what should be the algorithm of the neuron for this to happen, with low efficiency of the neuron, I will explain below.</p><p>The paradox of neuron efficiency can be explained by the fact that with an increase in the number of neurons in the nervous system, the “attention” of evolutionary processes to individual neurons decreases. Therefore, the neurons of the worm, one might say, work like a clock, very accurately, but the neurons of the grape snail and humans cannot boast of such accuracy.</p><p>So, two theories can be presented for the complex activity of a neuron: a neuron is a complex calculator whose operation algorithm is difficult to understand and justify, or a neuron simply works very unstably, which is compensated by its excessive amount, which is the simplest solution from the point of view of evolution. Apply to these theories Occam’s razor rule, according to which you need to leave ideas that have the simplest explanation, and most likely these ideas will be correct.</p><p>On the one hand, the paradox of neuron efficiency gives us a positive hope that the required computing power for brain modeling will require significantly less than a direct estimate by the number of neurons and synapses in the human brain. On the other hand, this greatly complicates the study of the biological brain. We can create a sufficiently detailed model of a small fragment of the cerebral cortex, spending a lot of computing power, and in this model, we cannot see any significant processes that would indicate how cognitive mechanisms in the nervous system proceed. Such attempts have already been made.</p><p>At first glance, the easiest way to create a general theory of how the brain works is to create a detailed model of the brain in accordance with the many scientific facts known about the neuron and synapses. Modeling is the most practical scientific tool for studying any complex system. The model literally reveals the essence of the object under study, allows you to influence the internal processes occurring in the modeled system, making it possible to understand them deeper.</p><p>The neuron does not have any exceptional organelles that would perform calculations, but its membrane has a number of features and allows the neuron to perform certain work. This work can be defined using a system of equations called the Hodgkin-Huxley model, which was developed in 1952, for which its authors received the Nobel Prize.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*0N2aHQAZL2ugHwFvNCnbrw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*QZKQcNmX56htMntXutKrew.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/360/1*8OXn96gtlcPC8zbT4O-CCw.gif" /></figure><p>These formulas contain several coefficients that determine some parameters of the neuron membrane, such as the reaction rate of ion channels, their conductivity, etc. This magical model describes several phenomena at once, in addition to the change in charge on the surface of the neuron membrane. First, it describes the neuron activation function, or summation mechanism, it is quite simple. If the initial charge is insufficient, then the model remains in equilibrium. If the charge crosses a certain threshold, then the model responds with one spike. If the charge significantly exceeds this threshold, then the model responds with a series of spikes. In computer neural networks, a wide variety of variants of the activation function are used, the closest to biology may be the Heaviside function (the unit step function) and the linear rectifier. But you need to understand that we are describing a fairly simple aspect of the operation of a neuron — summation. In my work on the tadpole mentioned above, I used a very simple version of the summation model, which can be figuratively represented as a vessel that accumulates a factor of stimulus, if this factor exceeded a certain threshold, then the neuron was activated. In order for this adder to work in real-time, the influence factor slowly expired from the vessel.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*LUcUGZAm_zfXtPdrQz5ntA.jpeg" /></figure><p>This summation model made it possible to summarize the signals that arrived at the neuron asynchronously, and it works quite realistically. In my opinion, the easier it is to describe this process, the better, and this is an essential difference between biological and computer networks.</p><p>Secondly, the Hodgkin-Huxley model describes the change in charge at one point of the membrane, but if, for example, we create a topologically accurate 3D model of a neuron and break this model into a uniform grid, we can apply the Hodgkin-Huxley model at each vertex (node) of this grid, subject to the influence of the charge on the value at adjacent vertices along the grid. Thus, we will obtain a model of the propagation of excitation along a neuron close to how it occurs in a living neuron.</p><p>The main conclusions that can be drawn from this model are that excitation, arising in any part of the membrane, spreads to the entire membrane, including propagating along the long axon to the most distant synapses. The Hodgkin-Huxley model is very resource-intensive, therefore, for modeling purposes, less expensive models with very similar graphs are used, several models have been invented.</p><p>As part of the Human Brain Project (HBP), a model of a small fragment of the mouse cerebral cortex was created, its creators took into account a lot. 3D models of neurons were recreated from real neurons, one of the variants of the Hodgkin-Huxley models was used, various types of neurons and neurotransmitters were taken into account, and there is no doubt that the model really corresponds to a biological analog. A lot of resources and time were spent on this, but it did not give significant results due to the fact that it was impossible to see significant processes in such a small size due to the paradox of neuron efficiency. Therefore, the path of repeating biology in detail is very, very time-consuming. The key to success is the ability to understand how nerve tissue and neurons work on a larger scale.</p><p>Let’s look at how the brain processes information on a particular example, on the processing of visual information. We will draw up a diagram of a neural network that performs this task.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C9vXsItDXn8wLRxKojOXUw.jpeg" /></figure><p>Information from the retina is transmitted through the optic nerve to the thalamus, where information is practically not subjected to significant transformations. Then it is transmitted to the primary visual area of the cerebral cortex (V1). Six layers are distinguished in the cerebral cortex, but these layers are histological or morphological. We are probably dealing with two layers here since some structures are repeated twice. But even then we are not dealing with two separate independent layers, layers of nerve cells working in tandem.</p><p>Let us characterize the area of the visual cortex V1 as the first layer in which information processing takes place. Zone V1 also has feedbacks with the thalamus. Similar feedbacks exist between all subsequent layers. These connections form cyclic transmissions of excitation between layers called reverberations.</p><p>After V1, information is transferred to the next zone V2, all subsequent zones will have less area. Depending on what the brain is observing, whether it was an object, a symbol, a person’s face, a place, or something else, information from V2 can be transmitted to different regions of V3, V4, V5. That is, already on this visual area V2, a serious categorization of visual images takes place. And approximately already on the third or fourth layer, it will be possible to select the neurons-detectors of certain images. For example, we can select a detector neuron for the letter “A”, number 3, or Jennifer Aniston’s face. By activating these detector neurons, we can judge what the brain is currently observing. A fairly simple architecture of a neural network, if we compare it with the architecture of computer neural networks specialized in the recognition of visual images, convolutional neural networks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I3hYuMknkm1a75TBU0-bGg.png" /><figcaption><em>AlexNet</em></figcaption></figure><p>There is something similar, this is a hierarchy of convolutional layers, each subsequent layer will have fewer and fewer parameters. But layers of this type of computer networks do not have feedbacks, of course, their presence is not a criterion for successful pattern recognition since the nature of reverberations in a living brain is not fully understood. There is a hypothesis that reverberations are associated with the phenomenon of instant memory, the memory that allows us, for example, not to get lost when dialing a phone number or pronouncing it. The reverberant activity is, as it were, delayed, denoting the areas through which this activity passes, thereby creating a context for the information being processed.</p><p>A person can recognize complex images in a fraction of a second, the speed of propagation of the action potential across the membrane is from 1 to 120 m / s, the synaptic delay in chemical synapses is 0.2–0.5 ms, which suggests that a chain of fewer than one hundred neurons.</p><p>The above suggests that in our cranium there is a neural network that works faster and more efficiently than any computer neural network, while it is organized relatively simply, performing simple transformations of information. This understanding encourages the search for an algorithm that would perform the task of pattern recognition using significantly less computing resources than modern neural networks.</p><h3>Artificial neuron</h3><p>Ever since my school years, I was worried about the idea of creating artificial intelligence, I satisfied my interest by studying the literature on neurophysiology, and I knew nothing about artificial neural networks. I got acquainted with neural networks later, as a student. Acquaintance with the artificial neuron McCulloch-Pits, which is the basis for all modern neural networks, puzzled and disappointed me, because of the great emphasis on dendritic synapses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1015/1*L2kDkbOszfl4WUzKZBlYHw.jpeg" /><figcaption>McCulloch-Pitts Neuron</figcaption></figure><p>The McCulloch-Pitts artificial neuron can be thought of as a function with many arguments and one response. The input arguments are multiplied with the appropriate coefficients, called weights (W1, W2, … Wn), then these values are added and the resulting sum passes through the activation function, the result of which is the result of the neuron’s calculations. The main thing is to choose the right weights, that is, train the neural network. This neuron model may seem simple and straightforward, but it has a strong emphasis on dendritic synapses.</p><p>In a chemical synapse, two important parts can be distinguished: they are presynapse and postsynapse. Presynapses are located at the ends of a long, single process of an axon, which can branch many times. Presynapse is presented as a small seal at the tips, it refers to the neuron that transmits excitation. Postsynapses are located on short branching dendrites, they belong to the neuron to which the excitation is transmitted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/886/1*-iy61hO6C_t0G7WCs7xFaQ.jpeg" /><figcaption>Synapse</figcaption></figure><p>In the presynapse are vesicles with portions of the neurotransmitter substance. It was in presynapses that the unequalness of synapses was previously revealed, presynapses differ in the number of portions of the neurotransmitter stored in it, as well as in the amount of neurotransmitter released when it is activated. The weight or strength of the presynapse is denoted by the letter S.</p><p>On the surface of the membrane of the postsynapse are receptors that respond to a neurotransmitter. The number of these receptors determines how sensitive the synapse is. That is, postsynapse can also be characterized by some characteristics, weight. Let’s designate this weight by the letter A. Of course, these two parameters can be represented as one W, which determines the strength of the entire synapse, but these parameters must be tuned in different ways during training and they still refer to different neurons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1Yp8PDarmsBifRh_O_miCA.jpeg" /><figcaption>Artificial neuron</figcaption></figure><p>Such a representation of a neuron is more realistic, but at the same time, it becomes much more complicated, since now it is necessary to understand how to adjust all these parameters during training.</p><p>I would like to present my version of the algorithm by which changes occur in postsynapses, that is, dendritic synapses. It is based on the fact that a biological neuron needs to maintain a certain level of activity. The fact is that a neuron as a cell is very resource-intensive for the body, it cannot feed on its own, it is for it that satellite cells, glia, do it. Therefore, if a neuron for some reason does not perform its functions, then the best option is to get rid of it in order to make the whole organism effective. With a prolonged absence of activation in a neuron, the process of apoptosis can start, this process is actively supported by satellite cells, literally tearing the neuron apart. Therefore, in order to survive, a neuron in conditions of an insufficient source of activation has to develop branches of dendrites, increase the sensitivity of synapses on dendrites and sometimes even migrate to other areas (this happens extremely rarely and under certain conditions), or produce spontaneous activity. This is evidenced, for example, by visual or auditory hallucinations in people whose organs of sight or hearing are deprived, or degraded due to aging. Oliver Sachs writes about this in more detail in his book The Man Who Mistook His Wife for a Hat.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FSgOTaXhbqPQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DSgOTaXhbqPQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FSgOTaXhbqPQ%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ae882a7e6cc7cbb9d24a5b6ade05ab25/href">https://medium.com/media/ae882a7e6cc7cbb9d24a5b6ade05ab25/href</a></iframe><p>On the other hand, excessive neuron activity can also lead to its death. The activity of a neuron is a very complex process that requires a clear implementation of many mechanisms, and any failure in their implementation will lead to fatal consequences for the entire cell. If the sources of activity are excessive, then neurons begin the process of degradation of some branches of the dendrites and a decrease in the sensitivity of their postsynapses. Thus, the neuron is trying to find some balance in the level of its activity, regulating dendritic synapses. The neuron, acting as an independent agent, acting in its own interests, provides amazing adaptability and plasticity of the entire brain. Despite the paradox of neuron efficiency, a healthy brain works very well, and each neuron plays a role. Due to this mechanism, neurons in the visual cortex of the cerebral cortex of blind people will be involved in other nervous processes not associated with the processing of visual images. And the excess in the number of nerve cells makes the nervous system very reliable, and if some parts of the nervous tissue are damaged, neurons can take over the functions and roles of the lost cells.</p><p>Based on this version, dendritic synapses are assigned a role that affects the adaptive qualities of the entire nervous system, and not some logical functions that determine cognitive processes.</p><p>For changes in the presynapses of axon synapses, an algorithm already exists, the so-called Hebb rule.</p><blockquote><em>When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A’s efficiency, as one of the cells firing B, is increased.</em></blockquote><blockquote>Hebb, D. O. The organization<br>of behavior: a neuropsychological theory. New York (2002)</blockquote><p>I am citing the full text of Hebb’s rule here because there are many interpretations of it that change its meaning.</p><p>As you can see, the focus is on the neuron that transmits the excitation, that is, on the axon synapses, and not on the dendritic synapses of the receiving neuron. Presynapse and postsynapse certainly influence each other. For example, with a deficit of activations, the neuron will first increase the sensitivity of the postsynapse that is more often used. And if it is necessary to reduce the level of activation, those postsynapses that were used least often will degrade first. This is due to the importance of preserving the logic of learning during adaptive processes.</p><p>If we want to create an artificial neural network, then we can neglect the adaptive mechanisms, after all, biological systems are more demanding to save resources with each element than artificial models.</p><p>It turns out that the basis of computer neural networks is a model of a neuron, in which the accents are placed the other way around than in a biological neuron. Therefore, one should not count on a high-quality result in the development of this direction. But understanding these problems, you can change the situation, you need to rebuild the concept of neural networks anew, revise it, laying the right foundation.</p><h3>Analysis and Synthesis</h3><p>Neurophysiology is a young, not yet mature science, it does not yet have strict fundamental laws like the laws in physics, although it contains a large number of theories and facts. It seems to me that the postulates and principles of the reflex theory of Ivan Petrovich Pavlov may be such laws. They can be compared to Newton’s laws in physics. When creating new theories in neurophysiology, we must ask questions: how reflexes occur and are formed within the framework of our theory, as well as how the processes of synthesis and analysis manifest themselves.</p><p>Analysis and synthesis require special attention. These concepts seem very abstract, but these are specific processes that take place in the nervous system. I.P. Pavlov believed that analysis and synthesis are continuous in the cerebral cortex. These processes are the basis for cognitive activity. I will try to convey in an accessible way what these processes are, it is very important in order to recreate cognitive processes in neural networks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*CTjhFco-trFkq2GHSH0ddg.jpeg" /></figure><p><strong>The synthesis is a mechanism for combining, generalizing various features into one image or action.</strong></p><p>An example from the experiments of I.P. Pavlova:<br>A specially trained model animal — a dog, isolated from other external stimuli during feeding, is turned on the sound of a metronome, which was previously indifferent to it. After several such combinations, the dog will develop a conditioned reflex, that is, to the sound of only the metronome, the model animal can produce gastric juice, as during feeding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*U66Lt8WOEon4A1Ca8T9cxA.jpeg" /></figure><p><strong>The analysis is a ranking mechanism (assigning the significance) of each feature from a limited set of features.</strong></p><p>An example from the works of I.P. Pavlova:<br>A previously trained model animal, in which a conditioned reflex to the sound of the metronome is formed, the production of gastric juice, the experimental conditions are changed, now the animals receive food at the sound of the metronome 120 beats per minute, and at the sound of 160 beats per minute, they will not be reinforced with anything. At first, the learned conditioned food reflex was triggered by both sounds of the metronome, but by the confluence of many repetitions, and much more times than in the experiment with synthesis. The dog begins to distinguish between these two very similar stimuli and stops responding to the sound of the metronome with a frequency that was not reinforced.</p><p>Let’s qualitatively compare these two cognitive processes.</p><p>The synthesis is a relatively fast mechanism because it requires a small number of examples; in turn, Analysis requires much larger iterations. The synthesis can proceed in some passive form, that is, the main thing here is the simultaneous combination of stimuli or signs so that they can be combined. Analysis always requires emotional reinforcement or some kind of feedback, which will determine which signs to increase or decrease their importance, rank. Synthesis always precedes Analysis, that is, the characteristics must first be combined into a group, within which ranking can already be performed (analysis process).</p><p>The analysis always leads to a reduction in the number of errors, since it gives the data additional information content: the ranks or significance of individual features. Pure synthesis creates a lot of errors, as it leads to a decrease in the information content of the initial data, combining them, generalizing them into single groups.</p><p>Now, armed with an understanding of these processes, let’s analyze computer neural networks for their presence.</p><p>Backpropagation of an error is pure analysis, it is the process of ranking the inputs of neurons based on the results of the entire neural network. There is no synthesis as a mechanism in neural networks. Each neuron has already formed a group of inputs from the beginning; this group does not change in any way in the process of learning according to the principle of synthesis. There may be a false impression of the presence of synthesis in neural networks due to their ability to classify data, but this is the result of the analysis mechanism working on the data. The synthesis is the ability to generalize, to merge data, and not to combine into groups based on common characteristics.</p><p>As a result, the high ability to generalize, which is inherent in human intelligence, is so badly lacking in computer neural networks, and this is offset by the need to use a large number of examples in training.</p><p>It should be understood that algorithms in which there is an emphasis on Analysis will still have an advantage in certain tasks. For example, in the task of searching for patterns in a large amount of data, or recognizing faces from a million database, no algorithm can compare with modern neural networks. But in a task where it is required to apply the experience gained from a small number of examples in various and diverse situations, for example, the autopilot task, then other new algorithms based on Synthesis and Analysis are required, similar to what happens in the brain.</p><h3>Instead of a conclusion</h3><p>What I am doing is looking for new algorithms, it is creating models based on the above principles. I am inspired by the study of the biological brain. This path goes through a series of failures and delusions, but with each new model, I gain valuable experience and become closer to my goal. In the future, I will try to share some of my experiences on the example of parsing specific models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*-Tj-ULKq5TRFsH-NVHrrjw.gif" /></figure><p>Now I set myself the task of creating a neural network algorithm that can distinguish handwritten numbers from the standard MNIST set, and no more than 1000 examples and presentations should be used during training. The result will be considered satisfactory with no more than a 5% error. I’m sure this is possible because our brains do something similar. Let me remind you that MNIST contains 6000 training examples, which can be presented several dozen times to set up a neural network.</p><p>Thanks!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ecd3c1fef57" width="1" height="1" alt=""><hr><p><a href="https://medium.com/good-audience/comparison-of-the-brain-with-a-computer-neural-network-3ecd3c1fef57">Comparison of the brain with a computer neural network</a> was originally published in <a href="https://medium.com/good-audience">Good Audience</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is the mind]]></title>
            <link>https://medium.datadriveninvestor.com/what-is-the-mind-b59d639b6c06?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b59d639b6c06</guid>
            <category><![CDATA[mind]]></category>
            <category><![CDATA[neurons]]></category>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Sun, 25 Nov 2018 20:43:32 GMT</pubDate>
            <atom:updated>2018-11-26T10:54:53.231Z</atom:updated>
            <content:encoded><![CDATA[<figure><a href="http://www.track.datadriveninvestor.com/1126B"><img alt="" src="https://cdn-images-1.medium.com/max/700/0*EXRTwovCWdH2FwtI" /></a></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZZb8byEJkeVIwf9k1cpWjA.jpeg" /></figure><p>One of the most important scientific questions of Mankind is the question: “What is the mind?”. How a person thinks, makes decisions, how thinking takes place, analysis and interpretation of various external stimuli, etc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*KkSeXVXKsZK_W9bNJcirJQ.jpeg" /><figcaption>Rene Descartes</figcaption></figure><p>Reflex is usually considered a response to an irritant and this concept is quite simple. It was introduced by Rene Descartes in the 17th century AD. Descartes presented the nervous system as a kind of hydraulic structure with “nervous tubes” that were filled with “animal spirits”, when they were applied to them they moved first to the brain, and then, reflecting also moving along the tubes, acted on the muscles, causing them to contract like a hydraulic executive mechanisms. The word reflex from Latin means reflected, and its essence is well reflected in the following scheme, which still retains its relevance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6NXZ_zKoaS54t-7UsCFWwA.jpeg" /></figure><p>Irritants act on receptors of sensory organs, receptors interpret these effects in nerve impulses, signals entering the central nervous system, the brain, where they are processed by appropriate neuronal chains (reflected) and then the corresponding reflex response occurs, muscle contraction or gland secretion.</p><p>But this scheme was not enough to explain many forms of targeted behavior. After all, it would be logical to say that if we stop sending sensory information, then stop nervous activity. For animals with a relatively simple nervous system, this is true, for example, if a frog cuts the ascending nerve pathways, then her brain seems to be plunged into a dream, and will not generate any nervous activity. But if the same is done with the cat, then there is the probability of finding the nervous activity leading, for example, to walking.</p><p>On a man, an operation to partially cut the spinal cord to test Descartes’ hypotheses was not done for ethical reasons, but the American psychologist Timothy Leary conducted experiments in special deprivation chambers. The Leary sensory deprivation chamber was a bath with a special saline solution, which kept the body of the experimental body afloat. The chamber was isolated from external sounds and light, the temperature of the solution was regulated and selected according to body temperature. American physicist Richard Phillips Feynman wrote in his autobiographical book, “You’re kidding, Mr. Feynman!” Usually Richard fell asleep, but it happened that he was experiencing some experience of out-of-body stay. In general, the complete cessation of nervous activity in the absence of irritants can not be said.</p><p>Hence, Descartes’ scheme is wrong and there is something that is above the reflex activity, a kind of thinking, or a thought process. — No! It is true, it just needs to be supplemented.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pFYKDYvrWnGZRCzHBxh9lQ.jpeg" /></figure><p>First, the Descartes scheme does not take into account the needs and emotional mechanisms. For example, food hunger can cause the activity of the corresponding “requirement cells” and their activity can lead to the activation of certain reflexes that would lead to targeted actions to meet the need for food. Our needs are a source of active action, which arises in the central nervous system itself. Nerve cells with receptors that react to leptin and its absence are in the hypothalamus. Leptin is produced by fat cells and is an indicator of the level of nutrients in the organism. Therefore, if we isolate the central nervous system from the body, activity in it will also arise because of the absence of leptin.</p><p>The need for novelty does not at all affect the periphery and the level of activity of the corresponding “requirement cells” depends on the nature of activity in the central nervous system itself, which creates a source of activity aimed at finding new information, learning new materials, reading books, or sliding one finger on the phone and tablet screen.</p><p>The activity of the “requirement cells” of unmet needs can be used — this is called sublimation. In particular, Sigmund Freud described sublimation as the redirection of sexual energy into a useful channel. Some creative individuals are credited with high creative productivity due to the use of nervous activity in scooped out of unmet needs or unbreakable pain.</p><p>Secondly, the nervous system has a unique property — memory. We know that this is not just information recorded somewhere, but a certain reorganization of the system. We can say that with each new external signal being processed, we are dealing with a new system. As Heraclitus would say: “You Cannot Step Into the Same River Twice”. Therefore, it is necessary to take into account the time factor when describing the nervous system. Reflex is not only a response to an irritant, it is a response to an irritant taking into account the entire history of stimuli received. For example, we have two twins of their nervous systems, in their structure are very similar, but in the process of growing up to one they addressed by name — Michael, to another — David. If we act on their nervous systems with the same stimulus, the question: “What’s your name?”, We get various responses, the answers: “Michael” or “David”, respectively. The history of the data obtained, both from external stimuli and from body signals, as well as the initial settings of the system determine what response to the stimulus will produce the nervous system at a given time.</p><p>Third, the generators. Generators play the main role in the processes of thinking. A generator is a chain of neurons in which a cyclic transmission of nervous excitation occurs. The generator, as it were, accumulates excitation in itself and can be its source. For example, the central generator of ordered activity, which provides rhythmic ordered motor signals without feedback.</p><p>The implementation of the generator in the simulator of the nervous system:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*Yacv9tmpzKZyxvd9svIfMA.gif" /></figure><p>Simple generator circuit:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4E5mtFXQEGrCFAMeV3I0fw.jpeg" /></figure><p>The generator is a closed circuit of a chain of neurons. The generator starts from the activation of the receptor “Q”, and its stopping by the “W” receptor through the inhibiting neuron, which inhibits the onset of excitation in one of the neurons in the chain. Such a closed loop can be a source of excitation. This neural network illustrates a simple reflex act, but at the same time during the activation of the generator some actions occur without stimuli.</p><p>What is it, a reflex or thinking? Some researchers prefer to call phenomena associated with the accumulation of nervous activity by thinking, but for me the term reflex is less abstract, it even implies the transfer of excitation from nerve cells to cells. Therefore, we will use the term reflex and reflex act, implying that the reflex does not always require an irritant and the generator can be a source of excitement.</p><p>Generators appeared in the nervous systems evolutionarily very early, mainly they are used to generate cyclic contractions of muscles during movement and performance of certain vegetative functions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7HzQuvLm16LonzefbR8jmQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*qUZ0avMybWg7lvLxlLJQLg.gif" /></figure><p>In a typical cockroach, there are two regimes of locomotion (movement in space): unhurried walk and running. When the cockroach is in search of food and studying the surrounding space, it slowly moves its legs one by one, and a chain of neurons — a generator that is activated by cells responsible for feeling hungry (Q) — can serve as a source of cyclic nervous activity for these actions. The disabling of this chain can be the inhibitory effect of the cockroach receptor analyzers, if the receptors indicate the presence of food, then one should not go further (W). In case of danger, the cockroach can choose a faster moving mode. The included light in the kitchen in the middle of the night will be an irritant to turn on the chain of neurons activating the run (R). In the above circuit, a modulated neuron (green) is used for switching, its activity indicates a state of panic or stress in the cockroach. After the end of the modulating action or the influence of other favorable factors, the uneven cockroach system switches back to the “leisurely walk” mode (F).</p><p>Of course, this scheme only demonstrates some principles of the organization of biological neural networks and is not an interpretation of the nervous system of the cockroach. The nervous system of the cockroach is much more complicated, it has hundreds of thousands of neurons and, of course, has a greater variability of behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2D1KTRncl1rKt-XEYlX5yQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*N1dBKKaNew9Rsj22Pie_uQ.gif" /></figure><p>Four-legged mammals increased the number of gait options compared to insects. In this example, the change in the types of gait occurs with the help of two stimuli “R” and “F”, and the transition from “Walk” to “Trot”, “Amble” and then “Gallop” occurs when the stimulus “R” is reactivated, order “F”, a kind of increase and decrease in transmission. Of course, such a complex action as walking can not be reduced to one-way signals sent to the extremities. Each limb is controlled by a group of extensor muscles and flexors, in turn, each muscle is divided into separate motor units and all need to give their agreed signals. There is also feedback, which is needed to correct commands in case of muscle fatigue or getting damaged. Theoretically, this model can be complicated to infinity approaching the biological analogue.</p><p>In addition to the programs of motor movements laid down in chains of neurons in mammals, there is a separate nerve center that allows to correct the work of the motor units — this is the cerebellum.</p><p>From the observation of the behavior of animals it clearly follows that in the nervous system there must be the possibility of switching between the various patterns of behavior embodied in the chains of neurons. For example, males of marsupial mouse (lat. Antechinus) during mating (once a year) radically change their behavior. Ignoring the need for food, water and without saving energy are in search of females or mated for 6 to 12 hours, and after that, having lost all the strength, perish. This is possible due to the modulating properties of neurons and synapses. The effect of a certain combination of hormones had a modulating effect on the neuronal chain-switch in the nervous system of the mouse, so the mouse reacted to the previous stimuli differently, in this case completely ignored its other needs except as a need for reproduction.</p><p>Modulation works even when we change our mood and when we manage concentration and attention. If your nervous system is modulated by the activity of dopamine neurons in the ventral tegmental area, the mood will be positive and will be able to enjoy life, move and learn something new, otherwise you will not even want to move.</p><p>The reflex becomes an increasingly complex concept: generators, memory, needs and switches — is it really only such a simple thing that can stand behind the greatness of Human thinking, consciousness capable of knowing the surrounding world and itself and its place in this world.</p><p>To dive further in response to the most important question, we consider the mechanisms of the nervous system with mental images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2kw1qcxuMspVMmCATf1CLw.jpeg" /></figure><p>The image in the nervous system is the activity of well-defined neurons, in the article on memory we saw how self-organization occurs and the specialization of nerve cells, which is based on the interaction of excitation of nervous tissue. The role that a neuron chooses for itself is determined by its location and the history of the processed data. The position of the sources of excitation relative to the nerve cell is important, they will be a reflection of the concept of “sign”.</p><p>The sign or set of signs after processing leads to the formation or activation of the image. The image in turn can generate a signal that will be a sign leading to the launch of another image. Consequently, we can talk about a certain hierarchy of images and we can identify levels of imagery or abstractness. Each subsequent level of abstractness will be less and less tied to the activity of certain receptors of the organs of perception. It is possible to compare the levels of abstractness with the division of various areas of the cortex as image analyzers of varying complexity.</p><p>The signs can belong to several different images at once, and the decisive condition for the activation of a particular image is an exceptional combination of signs, taking into account the mutual competition of similar images. Often, the formation of an image requires the participation of signs of a different nature, for example, for the visual recognition of objects by our brain, we need to use commands sent to the muscles that control the position of the eyes, as signs, along with some simple images.</p><p>People who are engaged in drawing, know that it is very difficult to draw a portrait without disturbing the position of the elements of the face. We see a face that we draw as a collection of separate images: lips, corners of lips, eyes, nose. And the face that is drawn by us is also perceived. Therefore, in drawing schools, it is recommended to initially create a certain outline, a skeleton of thin lines that will determine the position of the elements of the face.</p><p>There are brain damage in which people cannot recognize objects entirely, perceiving only individual signs. In the book “The Man Who Took His Wife for a Hat” Oliver Sacks, an American neurologist and writer, describes similar pathologies.</p><p>The combination of features as images of facial elements, and the size of saccades, as an estimate of the distance and position of these elements, allows us to explain why we easily recognize distorted or caricature faces.</p><p>Extend your hand in front of you and put your thumb up, an area with a diameter no larger than your thumb corresponds to the visual area that is perceived quite clearly for our visual system, the rest of the periphery, you can say, is perceived with low clarity, blurry. But we feel that the visual area of ​​our perception is much wider, this happens not only due to saccades, but the ability of our brain, in particular the visual analyzer, to “glue” the perceived images. Disruption in the work of this brain function makes people in life almost blind.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RcDni5kqgHj8hhN1bcKAOw.jpeg" /></figure><p>24 frames per second is the speed of a standard film projection, due to the inability of our perception to catch a frame shift at that speed. Our brain processes information not in discrete chunks, but in a continuous stream. Provided that for each stage of information processing it is required to spend a certain time, a situation arises in which information flows of varying degrees of relevance can be processed together. For example, the area called V5 (MT) of the mid-temporal cortex receives information in the form of some signs directly from the three areas V1, V2 and V3 of the primary visual information processing, which is processed in these areas sequentially. Accordingly, the most relevant information falling into the mid-temporal cortex is the information that comes from the V1 area, and the information from the V2 and V3 areas was relevant some time earlier. If information comes from three areas of the brain, it will differ in a certain way and there was no command to the saccade, then we can talk about the possible movement of the object, but if the eye movement was, then we can get an idea of ​​the shape of the object. Also, the V5 area can be used to evaluate the novelty of visual information; if the information is static in those areas, then it is time to make a new eye movement.</p><p>Our system of visual perception is based on several methods working in parallel, some methods are quick but inaccurate, others are better, but take time to collect signs in the form of concentrating on the details of an object.</p><p>The example with the visual area V5 shows how the brain can work with the context, but here we are talking only about fractions of seconds while the excitement spreads through the cortex. In the nervous system, there is a very simple mechanism that allows you to leave a “track” of information that has been processed, for use in subsequent processing as a context.</p><p>It is tempting to think that information processing in the nervous system proceeds sequentially from area to area and the signal does not return to its source, but in reality the structure and structure of the brain indicates the opposite. For example, all sensory paths pass through the thalamus, through the talamic nuclei, before reaching the cortex, almost all cells (90%) of the thalamus sending a signal to the cortex receive a reverse signal. And this tendency is characteristic of the whole brain, the V1 visual analyzer is inversely related to the V2 area and so on along the hierarchy. This gave rise to the theory of pulse reverberation as a temporary memory mechanism. In my opinion, it is true only in part. Generators can be elements of instant memory, the memory that is required when performing elementary actions, such as dialing a phone while we hear it. Instant memory lasts from several seconds to several minutes, with the longest reverberations in the prefrontal cortex, up to several minutes, and the reverberations between the thalamus and cortex areas analyzing sensory information last a fraction of a second or second, higher in abstraction levels, this time will increase. Total reverberations create brain rhythms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*TiyQPB8sQ3oqSeNCE-mT3A.gif" /></figure><p>The main thing that the generator works as a memory cell is the presence of lateral inhibition. Lateral inhibition is another mechanism that is widespread in the nervous system, from the retina and other sensory systems to the ganglia and the cortex. This system allows us to see sharper and sharper, select important sounds from noise and not confuse images. An example of the four generator elements is shown on the gif above; the operation of each generator suppresses activity in the other three. As you can see, this works fine, and there are no changes in the synapses and in general in the network structure, but we can say for sure which of the four signals was activated last.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f8Ip41qGYGdIx-8GUfAJbg.jpeg" /></figure><p>Now imagine a sensor map on the cortex divided into cortical columns, each column acting on its neighbors by lateral inhibition. This bark receives a complex pattern of activity from the receptor field of the sense organ through the thalamic nucleus, reverberation occurs, during which this pattern is modified. Weak and atypical signals are suppressed, and a more typical form of the image is formed for a given combination of features; this can be compared with how calculations take place in recurrent networks, but somewhat simpler.</p><p>The pattern of active excitation circuits will be fairly stable if subsequent signals from the receptor field differ slightly. It’s amazing how interconnected everything is in the nervous system, one mechanism is interlaced with another, and a memory element can also be an element of information processing. And only a comprehensive representation of the entire system gives a more accurate meaning of its individual mechanisms.</p><p>Another very important transmission circuit in the nervous system is the Peipets Circle (the anterior nucleus of the thalamus — the cingulate gyrus — the hippocampus — again the thalamus), this circuit closely interacts with the emotional centers of the rest of the limbic system. A distinctive feature is the hippocampus, in which the largest concentration of neurons with long-term potentiation is found. Long-term potentiation is an increase in the efficiency of synaptic transmission between neurons for a while from a few minutes, hours, or even days. This enhancement occurs due to the release of magnesium plugs from certain receptors on the postsynaptic membrane, so that this happens, repeated frequent passage of the action potential along the neuron membrane is necessary. We can say that by engaging in the reverberation of a hippocampal neuron becomes more sensitive, and it requires less impact to activate for a certain time. In turn, the neurons of the cingulate gyrus, like the other bulk of the neurons, are subject to the principles of habituation with prolonged frequent activation of their sensitivity decreases for some time. If one circuit element stops responding to a signal, then the reverb stops. The hippocampus is responsible for temporarily explicit memory, which is reflected in the long-term potentiation of its neurons. We use this memory throughout the day, and in the process of sleep an additional “run-through” of excitation transfer circuits marked with long-term potentiation occurs.</p><p>The Peipets circle is in close cooperation with the emotional centers, these centers determine which information the hippocampus will react to, modulating the sensitivity of its neurons.</p><p>Apparently, in the course of evolution, the central generator of ordered activity became more complex, more and more circuits of excitation transfer were added, conditions for interaction between the generators were added, branchings were added and the perimeter was increased. Thus, the crown of nature’s creation was formed — the human brain. As before, the described processes can be called reflex activity, although the architecture of the reflex has become much more complicated, but still it can be described systematically and theoretically modeled.</p><p>There is a generator unique to the human brain — this is a speech circle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bbPFccc-3YJ3g0576uLsDw.jpeg" /></figure><p>The speech circle is a circuit for transmitting information from hearing sensors and intramuscular sensitivity to areas of speech analyzers of the cerebral cortex, then to areas of speech reproduction, further to the muscles of the speech apparatus, and in turn the work of the speech apparatus activates certain sensory systems, and during the circulation of information there is a constant modification.</p><p>Simplified option:</p><p>Hearing (1), Intramuscular sensitivity (5)&gt; Wernicke (8)&gt; Broca (9)&gt; Muscle Activity (5)</p><p>In the process of pronouncing speech aloud, two sensory systems are activated — this is hearing (1) and intramuscular sensitivity of the muscles of the speech apparatus (5). Moreover, these two systems actually interpret one and the same information synchronously.</p><p>All sensory information passes through a region in the brain — the thalamus (2). The thalamus is a collection of nerve nodes or thalamic nuclei, which are groups and clusters of neurons. The human thalamus is a symmetrical formation having from 40 to 60 nuclei. The thalamus does not simply transmit information further to higher areas of the brain, but plays an important role in concentration, it is like a gatekeeper, stands at the entrance of the flow of information and assesses what to skip from the information and what to ignore. The nervous system is so arranged that it only works correctly at a certain level of brain activity, therefore mutual inhibition works between the thalamic nuclei, which forms a concentration mechanism. Concentration, for example, on hearing can suppress tactile sensations. Attention in the thalamus is regulated in two ways: “bottom up” and “top down”. The path “bottom up” is laid in animals from birth, We are inevitably attracted by loud sounds, new sounds, pain, unpleasant smells, etc. These signals are associated with reflex acts that increase (modulate) the sensitivity of the corresponding thalamic nuclei. The path “top bottom” is often a weaker control of attention and is carried out from the prefrontal cortex (10). We can, through our desires and our will, concentrate on certain senses and even on a certain part of the skin, but at the same time a loud sound will still shift our attention. Of course, everything is amenable to training and there are practices that allow us to develop attention management.</p><p>Already at the level of the thalamus information can be given an emotional assessment, for example, a loud and unexpected sound can activate the amygdala and cause a feeling of fear. Instinctively, the crying of a child is unpleasant to us, and the laughter of a child certainly evokes a feeling of joy.</p><p>After the thalamus (2), the information paths are distributed between the corresponding areas, the representatives of the cortex, information from the hearing organs enters the auditory cortex (3), and from the intramuscular receptors into the sensory cortex (6). In these areas, images of primary levels of abstraction are formed, then parts of these images merge into the associative cortex (4) and copies of the images of commands from the motor cortex (7) are added to the muscles of the vocal apparatus, all of these images will be signs for transferred to the Wernicke area (8).</p><p>Wernicke’s area (8) is responsible for speech perception. A person with damage to the Wernicke area can have excellent hearing and recognize and distinguish different sounds, but is not able to understand speech, including his own. As mentioned earlier, two sensory systems: hearing and intramuscular sensitivity synchronously form images that interpret one and the same information, but the total image from the two systems is perceived by Wernicke’s field, more precisely three should be added and copies of commands from the motor cortex to the muscles of the vocal apparatus. If the sensory information from the hearing is stopped, and only the sensitivity of the muscles remains, then Wernicke’s field will still “hear” this speech, the associative connection of these obtained images is very strong and for the associative cortex it doesn’t matter what features will form the image.</p><p>A person constantly conducts an internal monologue, his peculiarity is that the muscles of the speech apparatus make very weak contractions, which do not lead to the uttering of sounds and generally visible movements, but are sufficient to fix these contractions with intramuscular receptors. Reverberations arise between the Wernicke field (8) and the associative crust (4), which provide some context of information and associative links.</p><p>Images of Wernicke’s field (8) as signs are transmitted to Broca’s area (9) by means of an arc beam — a nerve cluster. Broca’s field (9) — the area of ​​the cerebral cortex responsible for speech reproduction. When the Broca area is damaged, a person can perfectly understand someone else’s speech, but when trying to speak instead of speech, inarticulate sounds are reproduced, or it is possible to play only one word. But Broca’s field is also important in the perception of sound, which is reflected in the severe defeat of the area. Contour: Broca’s field (9), motor cortex (7), associative cortex (4), and Wernicke’s field (8) are important for the formation of chains of sounds that form words; in turn, chains of words form phrases and sentences.</p><p>During speech, Broca’s field (9) is involved in reverberations with the prefrontal cortex (10). The prefrotal cortex (10) is a very extensive area of ​​the cerebral cortex, it is she who is responsible for the comprehension of what is happening at a given point in time. Reverberations with the participation of the pre-frontal cortex and in it themselves determine the instant memory, the memory of information that is necessary in the process of performing specific actions while we keep them in our field of attention, for about a few minutes. In addition, our prefrontal cortex (10) can have an inhibitory effect on emotional centers, thereby reducing their impact on our behavior.</p><p>Damage to the prefrontal cortex (10) makes a person more impulsive, makes him subject to vices, and actions become less deliberate and reasonable. It can be said that only due to the constant activity of the prefrontal cortex we do not obey the first call of our needs, for example, the desire to empty the bladder while in an important meeting, but allow us to see it through and do everything in the right place. Management of emotional centers allows you to determine what information will be stored and processed in the circle of Peipets (13). Please note that the information from the thalamus (2) falls not only into the analyzer areas, but also interesting and useful information is stored in the belt gyrus here for a longer time.</p><p>The main contour of the speech circle in the internal speech begins with the muscles of the speech apparatus (5), then the thalamus (2), sensory cortex (6), associative cortex (4), Wernicke’s field (8), then Broca’s field (9). Broca, in turn, “communicates” with the prefrontal cortex (10) and sends commands to the motor cortex (7). The motor cortex (7) sends commands to the basal ganglia (11) and a copy of these commands (12) to the cerebellum. The cerebellum (12) corrects commands of the motor cortex, dividing the work of motor units more harmonious and coordinated in time. Damage to the cerebellum can lead to slower speech, as the formation of motor action becomes more difficult. The basal ganglia form the final form of the command for the muscles of the vocal apparatus (5).</p><p>Very complex generator.</p><p>It is important to note that in the speech circle with the internal speech there is a “physical basis” — muscular activity. This makes internal speech subject to control, for example, during sleep a decrease in the tone level of all muscles, which deprives the internal monologue of feedback through muscle sensitivity, therefore only a small contour is possible (8, 9, 7, 4, 8). Unbeknownst to the prefrontal cortex (10) in the process of sleep, when the inhibitory effect on the emotional centers decreases, the Peipets circle (13) is activated and triggers images that could cause an increased emotional evaluation during the day, this is what creates the dreams. In his work “The Interpretation of Dreams” Sigmund Freud very successfully and accurately described the principle of dreams. The basis of visible dreams is a simple phrase or sentence, which has for us a significant value at the time of falling asleep, but we do not hear it, but only see visual images interpreted on its basis. Not infrequently, without additional control with the help of a “physical basis” a phrase can turn into nonsense.</p><p>Speech is a tool that allows us to transmit and accumulate information, declare and plan actions and events, thanks to this tool, Man was able to create a civilization. The main form of our thinking is internal speech, internal monologue and most of the time in this monologue we devote to social interactions, working through upcoming dialogues, or imaginary dialogues, for example, we go home from work and can predict what we will say to our wife (spouse, mom , brother, friend) at the meeting, that she (he) answer that we will answer in response. And these banal and mundane things are occupied by our mind constantly, if you are not a philosopher-thinker, who soars in the clouds. Speech allows us to declare all aspects of our life in a certain system of signs and meanings. Developed speech is a distinctive feature of Man from animals.</p><p>Now, having an idea of how the brain works and how speech is formed in this brain, we can answer the question: “What is mind and where is it localized?”.</p><p>Our nervous system is a single whole mechanism that can be divided into separate functional parts. Select individual chains of neurons, neural networks performing a specific task, for example, you can select sensory analyzers or, as in the example above, the contour chains are responsible for speech. I call these functional neural networks “personalities”, as much indicates their certain independence. Typically, these personalities in a healthy nervous system exchange information, they inform each other about what they are doing at the moment, about what is happening at the moment. This is due to the large number of connections between areas of the brain. Personalities act in cooperation, as if this is a coordinated team, no one is trying to go against the team. And the reason is simply associative learning, the events taking place at the same time will unite.</p><p>Schemes formed by neurons and configurations of connections between them in the Human nervous system can be divided into very small and simple ones, but we will select only a few basic ones. First, it is possible to distinguish sensory analyzers, for different types of sensory signals, their structures are defined. For visual information, these are thalamic visuals nucleus and the occipital portion of the cerebral cortex. Rumor — areas in the temporal lobes, sensory information — these are the parietal areas of the cortex, the taste is a small area in the insular lobes, the sense of smell is the olfactory bulbs and a small area in the temporal areas. The task of these neural networks is the primary processing of sensory information, as a result of the work, the formation of a certain image and its transfer to the associative regions of the cortex. The associative cortex connects various images, from analyzers forming their images based on them, this area is responsible for the perception of the surrounding world, it is this that forms the integrity of the picture of the world surrounding us. There are also chains of neurons capable of describing, declaring images formed by the associative cortex, they are localized in the Broca and Wernicke zones, but speech mechanisms may extend beyond these areas. The area responsible for making decisions is the prefrontal cortex, our thoughts in the form of inner speech are reverberations between the prefrontal cortex and the Broca area. Further, it is possible to isolate the areas responsible for motor actions, these areas are subject to a certain hierarchy, at the head of the motor cortex, then the basal ganglia and the cerebellum, and the formation of basic movements is possible in the reticular formation and spinal cord. But at the same time, neural networks in these formations can be completely independent, if there are no commands from above.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VpciFazSlmWo5sxlzJ_-bw.jpeg" /></figure><p>Like electrical circuits that can perform their functions only in the presence of electricity, biological neural networks operate only in the presence of nervous excitation. And here is the most interesting! This nervous excitation is limited for neural circuits. It so happened that the brain works correctly only when a certain level of activity of the nervous tissue is maintained, increased activity can lead to an epileptic seizure, a reduced one can plunge the brain into sleep. The level of activity is regulated by the thalamus, through the filter of incoming signals, as well as by the mechanism of lateral inhibition / inducement in the cortex and is maintained at a certain level (the range is quite wide). Thus, neural activity in the brain is a resource that is distributed among neural circuits. Of course, this distribution is uneven and depends on functional features, for example, visual analyzers require a large part of the nervous excitation as a resource to process a large amount of data from receptors. Distribute nervous excitement among the prefrontal cortex, the associative cortex and the areas responsible for speech — this is what our “I” is, our consciousness, our mind.</p><p>Rene Descartes was looking for at least one structure of the brain that would be unpaired, and as a result decided that it was a small pineal gland (epiphysis) behind the brain stem — is a container of the soul, because he considered the soul to be something indivisible. Many psychologists and philosophers since ancient times consider consciousness to be integral, indivisible and something permanent. After all, our brain creates this illusion extremely convincingly. But it turned out that the epiphysis consists of two symmetrical halves, almost mirrored in relation to each other.</p><p>The mind is the essence of the unbelievable fragmentation, up to the activity of a single neuron. In some cases, nervous activity can be concentrated in certain areas of the brain, for example, when solving a complex logical problem, emphasis is needed on the prefrontal cortex, and when relaxed inaction, the excitation is dispersed throughout the cortex. When we go to describe our subjective sensations, the focus of perception “shifts” to the speech areas, thus we can perfectly describe only the “declarative self”. When the focus of perception is strongly shifted to the prefrontal cortex, the consciousness leaves the zone of speech description. It can be calculated that the mind is something that obeys an internal monologue, but we make many decisions without the participation of internal speech.</p><p>What is the nature of the mind help to understand the various kinds of pathology in the brain.</p><p>Callosotomy is an operation to dissect the corpus callosum, a region of the brain that is an accumulation of nerve paths connecting the two hemispheres of the brain, allowing them to exchange information. This operation is performed to reduce the effects of epileptic seizures. One of the side effects of the operation is the syndrome of someone else’s hand. In the human body, it is as if there are two personalities who control the different halves of the body. Since the main speech centers are enclosed in the left hemisphere (in most cases), the person controlling the right hand will answer the questions. This person may complain about inconsistent actions from the left hand, which can take things, or perform actions that were not included in the plans of the left hemisphere. This occurs due to the fact that the “personalities” ceased to hear each other, which leads to mismatched work of neural networks.</p><p>Another example. Apraxia is a neurological condition characterized by an inability to perform targeted movements, despite the fact that a person knows what is required of him, wants to do it and possesses the necessary physical data. Apraxia can occur due to the rupture of communication between the areas responsible for the formation of motor commands and the area responsible for making decisions, the prefrontal cortex.</p><p>Breaking the connection between the visual analyzer and the prefrontal cortex can lead to seeing blindness or cortical blindness, in which the person is blind, but at the same time able to bypass obstacles, or with a high probability to guess which way the target-point moves during the experiments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DbNJNppmBEwjd_-WeykshQ.jpeg" /></figure><p>I recommend watching Jill Bolte Taylor speak at a TED conference. In the context of the above, the situation described to her becomes clearer.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fembed.ted.com%2Ftalks%2Fjill_bolte_taylor_s_powerful_stroke_of_insight&amp;url=https%3A%2F%2Fwww.ted.com%2Ftalks%2Fjill_bolte_taylor_s_powerful_stroke_of_insight&amp;image=https%3A%2F%2Fpi.tedcdn.com%2Fr%2Fpe.tedcdn.com%2Fimages%2Fted%2Feefe30d20338d800bdc70a09dc0f6007e7355a74_2880x1620.jpg%3Fh%3D316%26w%3D560&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=ted" width="560" height="316" frameborder="0" scrolling="no"><a href="https://medium.com/media/d98aa80d900fa19ae5325ee079b80f79/href">https://medium.com/media/d98aa80d900fa19ae5325ee079b80f79/href</a></iframe><p>Summing up, we can say that the answer to the question about the mind has long been disclosed, it remains only to accept it. Reflex activity, the speech circle, reverberations and the circulation of nervous excitement have long been known phenomena, but it is very tempting to create a mystical aura around the concept of mind and to search endlessly for some miraculous interpretation of it.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2Fb2a0d6%3Fas_embed%3Dtrue&amp;dntp=1&amp;display_name=Upscribe&amp;url=https%3A%2F%2Fupscri.be%2Fb2a0d6%2F&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/0707f5c806284d01a4a13c7b13a91ce3/href">https://medium.com/media/0707f5c806284d01a4a13c7b13a91ce3/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b59d639b6c06" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/what-is-the-mind-b59d639b6c06">What is the mind</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modeling of the cerebellum]]></title>
            <link>https://medium.datadriveninvestor.com/modeling-of-the-cerebellum-c841023a4dcf?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c841023a4dcf</guid>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[cerebellum]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Mon, 22 Oct 2018 17:21:59 GMT</pubDate>
            <atom:updated>2018-10-22T17:21:59.858Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UBEYcU12kG7gRRcaH3ZB1Q.jpeg" /></figure><p>What is surprising about the cerebellum? This is the department of the brain, which gave us the opportunity to enjoy the beauty and splendor of music, incredible accuracy and grace of dance movements. This is an amazing department of the brain that performs a very meaningful function.</p><p>The brain and the nervous system as a whole work with the same discrete signals, short pulses, the amplitude of which is the same. The way to control different organs is to change the number of pulses per unit of time, that is, the frequency of the signal. Also, the degree of influence on the receptor is determined precisely by the change in the pulse frequency.</p><p>Our motor system has the potential to very precise actions, to smooth movements or accurate movements of our limbs. To achieve this, muscles that require high flexibility in management have been divided into motor units. The motor unit is a motor neuron and the associated muscle fibers. And the more the motor units are involved in the activity, the greater the force is due to the contraction of the muscle. Also the muscle fibers of different motor units are intertwined for the purpose of load distribution and field of action. Thus, it is possible to control the muscles not only due to the frequency of activation of the motor units, but the total number of them involved in the activity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/846/1*v-M49nvLiezpR97jnfOqPw.gif" /></figure><p>Imagine a simple reflex, for which activation of its header leads to a complete contraction of the muscle. In such a reflex, all motor units will be activated almost at the same time. But in the process of vital activity, there may be a need, for example, to smoothly reduce, or plan to relax the muscle, or it requires only a smooth relaxation of the muscle, but then a rapid contraction of the muscle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*B_IiILpZQveVom_cAGumWg.gif" /></figure><p>In the simulated example, several reflexes lead to a different character of activation of the hypothetical muscle, taking into account that each indicator of action is the motor unit of one muscle. In the example, there is a basic reflex, activation of which leads to the simultaneous launch of all motor units. On the basis of this reflex, reflexes are created that ensure the sequential switching on or off of the motor units, which adds a certain time factor.</p><p>In the process of development, the animal learns to control its body, and learning occurs not only in the form of creating associations, linking certain events, but also the ratio of these events in time, and quite accurately. Of course, the propagation of excitation through the reflex arc gives a time factor, but the mechanisms of the formation of reflexes do not allow making this factor sufficiently accurate. In the nervous system, there should be a mechanism that allows you to remember the time intervals between actions, and it is flexible enough for the possibility of retraining.</p><p>When developing the theory of brain work and studying the nervous system, I was very surprised by the cerebellum, its features are very different from the rest of the brain structures. First, this is the ratio of the size of the cerebellum and the prescribed functional. The cerebellum occupies approximately 14% of the entire brain (for animals the ratio will be greater) and, as is known, serves to coordinate our body, and also performs some autonomic functions. In this case, the defeat of the cerebellum causes a violation of coordination in the form of the difficulty of maintaining balance, accuracy of movements, uneven handwriting, slow speech, etc. For the success of the animal, these functions are very important, but the size of the cerebellum is larger, for example, than the Wernicke and Broca areas together, although these areas are responsible for human speech, that is, they perform more complex functionality with smaller dimensions.</p><p>Secondly, there are unusual neurons in the cerebellum — the Purkinje cells. The dendrites of these cells strongly branch from one side of the cell nucleus, branching in one plane. Such a form of the cell must be something caused, usually the dendrites of neurons grow without strict orientation in the planes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/369/1*BY13DSth5L6r7swc-2ZR5w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/356/1*UWJ0z01XHElPckA7Z1JVeQ.gif" /></figure><p>Third, the structure of the cerebellar cortex differs significantly from the structure of the cortex of the neocortex. In the cerebellar cortex, only three layers are distinguished: the molecular layer, the Purkinje cell layer and the granular layer. If the evolution of the neocortex increased the number of its layers, then this did not happen in the cerebellum. Usually evolution does not touch what works well and sufficiently performs its functional. Given that the cerebellum evolves quite early, we can say that the structure of its cortex is perfectly refined by evolution under its functions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3XUpdfuAo6pmvLKfqa0wHA.jpeg" /></figure><p>One of the features of the structure of the cerebellar cortex is the presence of parallel fibers — these are the axons of granular cells that propagate parallel to each other and perpendicular to the plane of growth of the dur- drites of Purkinje cells. Probably, the shape of Purkinje cells is due to the fact that it increases the probability of the contact of the parallel fiber and the dendrites of the Purkinje cell with their growth and development of the cortex. The Purkinje cell is a “trap” for parallel fibers of granular cells.</p><p>It is known that the action potential has the form of a short pulse and propagates through the nerve tissue with a limited velocity, approximating 1m/s. This leads to the idea that the complex of parallel fibers and Purkinje cells is a kind of timer that is triggered by activating the granular cell, and the time intervals are counted by the sequential activation of Purkinje cells by the propagated action potential along the parallel fibers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*cmW7827abRdTEOELh3flvw.jpeg" /></figure><p>Based on this assumption, a model was developed that represents a strong simplification of the structure of the cerebellum. Usually Purkinje cells have an inhibitory effect in the nuclei of the cerebellum, and many other discrepancies, but we will not take all this into account, for the sake of simplicity. The aim of the model is to form an understanding of the principles of the cerebellum.</p><p>Imagine a simple reflex, an action will take place on a single stimulus. And we need to train the neural network so that the given stimulus was not a single answer, but a series of responses with a certain rhythm. It is almost impossible to achieve this by the usual associative method. You need some kind of tool that counts the time between events. Such a tool is the timing expansion, it is an analog of the biological cerebellum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/504/1*b4gIy-YCfhJihRexQQfgpQ.jpeg" /></figure><p>At the heart of the timing expansion is a sequentially connected chain of neural elements. The signal that runs this chain passes through the collateral inhibition scheme, this circuit skips only a single signal, it is necessary that there would be no interaction and competition between the elements of the chain. All changes leading to learning occur in synapses of neurons with high neuroplasticity. Each such neuron has its own element in the scheme, which is an analog of the complex of Purkinje cells and parallel fibers.</p><p>When the activity of a neuron with high neuroplasticity coincides with a neuron by the representative of a part of the reflex arc, an increase in the strength and strengthening of the synapse connecting them takes place. In the absence of coincidence, there will be a decrease in strength. The rate of these changes depends on the measure of the neural plasticity of the neurons. Initially, the strength of synapses of plastic neurons is not enough to activate a representative, only after a few consecutive synapses the synapse action will be enough to activate. After a few repetitions, this position will be strengthened. Of course, we can control the speed of training by changing the degree of plasticity in neurons.</p><p>Now if we repeat the reflex, activating the stimulus with the necessary rhythm, then after a while, with a single activation of the stimulus, we get the answer in a memorized rhythm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*9pfunjQD2494xsp5ic1Xrw.gif" /></figure><p>In this example, rhythm is memorized. Of course, any rhythm can be set. <a href="https://medium.com/@it.belkin/simulator-of-the-nervous-system-b0b69b4f346a">The example scheme is based on model neurons that were described in earlier articles.</a></p><p>It can be concluded that the function of the cerebellum is the preservation in memory of short time intervals in various motor actions.</p><p>As can be seen from the examples of the work of the timing expansion, it can work only with one motor unit, for each motor unit it needs its own complex. Having for each motor unit its own chain of Purkinje cells, it is possible to learn not only to “beat out” the rhythms, but also to the nature of the movements (smoothness, sharpness) due to the coordination of each motor unit in time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/468/1*xtODWwB2LrF3uBmzZ3CElg.png" /></figure><p>There is an increase in cerebellum in people whose activities are related to tasks where it is required to keep in memory the exact time intervals (timings) in motor actions. It is, above all, musicians, dancers and athletes.</p><p>It can be noted that a significant number of cells are required to store in the memory of time intervals, this causes a large relative size of the cerebellum. But the scheme chosen by evolution is primarily aimed at reliability and flexibility.</p><p>When creating principle models of the brain, it is not necessary to take into account all the schemes selected by evolution and it is possible to use a model neuron working like a whole complex of Purkinje cells and parallel fibers.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c841023a4dcf" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/modeling-of-the-cerebellum-c841023a4dcf">Modeling of the cerebellum</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Simulation of emotions]]></title>
            <link>https://medium.datadriveninvestor.com/simulation-of-emotions-2e01b2dc2b2e?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2e01b2dc2b2e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Mon, 10 Sep 2018 17:54:24 GMT</pubDate>
            <atom:updated>2018-09-10T17:54:24.678Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JnOZv_BkSKNGMztOAwZHZQ.jpeg" /></figure><p>Without understanding the nature of the emergence of intelligence and the mechanisms of intelligent behavior, it is impossible to create artificial intelligence. It is possible to compile networks of neuron-like elements indefinitely by varying their number, organization, and algorithms, but it is impossible to repeat it randomly the way evolution has for millions of years. I’m not saying that it’s impossible to create an intellect based on some other fundamentals, distinct from the basics of the human intellect. To fly it is not necessary model aircraft based only on flapping wings, but it is necessary to know the laws of aerodynamics in order to design them. Therefore, first one needs to understand the workings of the only known example of intelligence on the planet Earth.</p><p>The nature of the activity of living organisms is determined by their needs, such as the need to survive, eat, multiply and conserve energy. No living being acts outside the inner motives formed by their needs, and it is impossible to imagine a mind devoid of inner motives and goals. Every movement of our eyes, utterance of words, and formation of thoughts is the result of the mechanisms responsible for satisfying our needs. The creation of an absolute intellect, not conditioned by needs, would be the creation of a lifeless automaton or mechanistic cog. Our surrounding world is very complex and diverse, and the creation of reflexes for all occasions that would allow the organism to be successful in this world is not possible. Therefore, nature has created a mechanism that allows us to form and select the reflexes that lead to the desired result — these are emotions, or emotional mechanisms.</p><p>But what kind of needs can an artificial intelligence have? Because it does not need to experience hunger, and a sense of fear can provoke actions to self-defense; love and affection — these feelings even for humans have create a lot of hassle. The only thing that is really necessary for artificial intelligence is the desire to learn, curiosity and the desire for novelty. The nature of all emotion is similar and all of the above emotions may be embodied in artificial intelligence.</p><p>In our brain, there are many different areas, and we can say that there are certain interrelations between these areas; these connections are of a higher level than those that exist between neurons. These departments influence each other through chemical signals, this is a slower type of information channel. Each emotion is conditioned by its own area in the brain, the causes of its origin, by the effects it has on other areas, and, as a consequence, on the behavior of the organism itself.</p><h4>Vital needs and emotional mechanisms associated with them</h4><p>For most emotions, two states can be distinguished: hunger and satiety. Thirst and quenching thirst, prolonged muscle tension and relaxation, fear and sense of avoiding danger, pain and retreat of pain, boredom and novel experience, lack of an object of attachment and the joy of contact with the object of attachment are the two opposite parts of emotional mechanisms.</p><p><strong><em>What is hunger?</em></strong></p><p>Imagine a certain area of neurons in the nervous system, in which the activity will increase as a need increases, the more the need increases, the more often the activation of these cells rises and the more they become involved in mental activity. And accordingly, if there is no need, these cells will not be activated. You can call these cells — requirements cells, they seem to require the implementation of the actions necessary for calming them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/841/1*8zJ6katFkZgXB2X8bGvfag.jpeg" /></figure><p>Try to hold your breath, and you will feel the action of the “requirement cells” of the respiratory center. Over time, with a delay in breathing, signals from internal organs about the increase in carbon dioxide will come to the nerve center responsible for breathing, increasing its activity. Activity increases so much that it is impossible to restrain breathing through the inhibitory effect of the cerebral cortex. And this nerve center by its activity will cause the necessary reflex for breath. For each need, there exist corresponding “requirement cells”, although in some cases the work of groups of these cells is interrelated.</p><p>A living organism can have many needs and there are priorities between these needs, the need for self-protection is always higher than the need to satisfy food hunger. These priorities are ordered by means of the maximum level of activity for the corresponding requirement cells and mutual suppression. The organism will perform those reflexes that are related to the satisfaction of the requirement cells having primacy in its nervous system at the present moment.</p><p>We will study how needs and emotions control the behavior of animals, and what internal processes take place with the help of a computer model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*7SHTSbZORd4IA7bihOVieA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*e6giW4mqfj6lMECDFMdh5Q.jpeg" /></figure><p>To do this, we will look at a specific system in the program. There are two areas in the system, A and B. These areas are necessary so that we can logically separate and isolate groups of neurons. Area A is responsible for some need, and area B is encoded with a set of simple reflexes. Three reflexes with the headings Q, W, E (a) whose activation leads to the three actions “1”, “2”, “3” respectively (b), each reflex has representations in the area A (с), the transmission of excitation between the reflex and the representation is bilateral. Such an organization is derived from the principles of the organization of the nervous system, but it has been greatly simplified. Areas responsible for emotions and needs are localized in the limbic system, these areas are closely related to the hippocampus, which is responsible for temporary memory and has many cortical representations. And with virtual access to the hippocampus, we have access to a variety of cortical reflexes.</p><p>Through stimulus A, we can activate the need (d), this stimulus can symbolize, for example, the signal from the internal organs about the lack of necessary nutrients, which leads to a feeling of hunger and the activation of the corresponding requirement cells (e). Also, there may be a number of additional stimuli that have only a modulating effect on the requirement cells, for example, appetizing smells of food can increase the sensitivity of cells to food starvation requirements, which will increase their activity. But with absolute absence of hunger, the smell of food will not affect the behavior of the animal.</p><p>Suppose that when action “2” is performed by animals, it leads to satisfaction of a certain need (f). For satiation, the neuron (g) will be responsible, the activity of which leads to imprinting in area A. In this case, the neuroplasticity changes for one second from 0 to 0.1 and back (h). When we discussed the topic of memory, we talked about the mechanism of emotional memory and this is an example of how this mechanism is used. In the same way, the center responsible for the activity of the requirement cells (i) is suppressed, but this suppression is delayed, and this gives rise to the possibility of forming associative links. An analogue of these processes is the activation of dopamine neurons. The dopamine mediator is inhibitory, it has a retarding effect in the areas responsible for hunger. Also, the release of various “happiness hormones” that affect the plasticity of some areas in the brain change the ability to concentrate attention and perception, and reduce the pain threshold, which in aggregate can be interpreted as a state of euphoria, or an uplift in mood.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*VOtQDioRVJ7mvVHOhM0hHw.gif" /></figure><p>Even with the activity of the “requirement cells”, actions that do not lead to satiation do not cause any significant changes in the region A. But at the first start of the reflex leading to satiation, a reflex arc is created that connects the “requirement cells” and the representation of the reflex leading to the desired actions. Thus, a certain closed contour is formed; the emergence of the need leads to the activation of “requirement cells”, which activates the reflex that forms the actions leading to activation of the satiation center. The feeling of satiation extinguishes activity in the “requirement cells”, until the need arises again. With each repetition, the formed reflex arc will be strengthened by increasing the strength of the synapses. In this case, the value of the maximum force of the synapse 100 is set, it will take 10 repetitions to reach the maximum value of the stability of the reflex act.</p><p>In the event that we further cancel the condition: by the action of “2” lead to saturation, then when the need arises, this action will be aggressively carried out, at first very actively, but over time, the habituation processes will lead to the attenuation of the performance of this reflex, and the adaptation processes will contribute to periodic return to action “2”. There are not any mechanisms for erasing information, and the reflex arc will persist. The reflex arc can be lost due to retraining and participation of neurons belonging to it in other reflex acts. In biological systems a situation is possible when useless actions are suppressed by the needs associated with the conservation of energy. In fact, the need to conserve energy will always arise when performing any actions, but is satiated when performing reflexes that suppress the action of motor neurons as a result of inaction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*e91KVnyOcINA0JU1VYdcDg.gif" /></figure><p>Another option is to add a condition such that action “1” will lead to satiation. In this case it is not enough just to introduce a new condition, it is necessary to train the organism at the very beginning that there is such an opportunity. After several acts of training, when a need arises, both actions will be performed, although these actions may be in competition when performed simultaneously. More often the reflex act will be performed, the reflex arc of which is more stable as there are stronger links, and in the end this reflex will completely dominate. Simply put, the organism is more likely to perform those actions that previously led to positive results.</p><p>If, for example, the execution of action “1” is more effective for saving energy, then it will give one an advantage over action “2”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*4lJIaWCaaRHQufnFkzp1fw.gif" /></figure><p>And of course it is possible to completely retrain, even if there is a maximum stable reflex, given that you establish the appearance of a sense of satiation only in response to the action of “1”. This will require some training. After that, for some time, action “2” (“out of habit”) may be performed, but in the end the organism will only perform action “1” when the need arises.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*S11Q9IBcTyUSyEwcA7Kmog.gif" /></figure><p>The considered relationship between emotions and the satisfaction of needs is possible only in the presence of some preconfigured reflexes and a teacher, which at the appropriate moment will lead to the implementation of the necessary actions. In fact, this is quite consistent with the natural world. But what about the situation in which an organism tries to study the environment on its own? This is possible due to the very important need for new information, or novelty seeking.</p><h4>The need for novelty</h4><p>The emotional mechanism behind the need for novelty works according to principles similar to those of any other need, and because of this we need to determine a certain criterion for the novelty of incoming information. If for some time the body does not receive new information, then the need for novelty will begin to arise and activation of the corresponding “requirement cells” begin to occur.</p><p>From an evolutionarily perspective the need for obtaining new information appears relatively early, and physiologists have found some basic neural networks for assessing the novelty of information in the nervous system. For example, a neural network that allows you to identify abrupt changes in sound level, which is necessary to attract the attention to changing aspects of the environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*yPMKZVN4AOturPKRh5w5Cw.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/448/1*jmBIgoBKc4UHPVoLNaffaA.jpeg" /></figure><p>This neural network responds only at the beginning of a series of spikes. If you have such networks for the whole spectrum of sounds, then you can only select new sounds and react to them.</p><p>But it is difficult to imagine a scheme that can evaluate the novelty of information at the level of abstract thinking, the level of formation of images and associations. For a model neuron, there is a very simple way to evaluate new information, the change in the cumulative state and changes in synapses. Having this criterion of novelty for each neuron at the moment of its activation, it is possible to obtain a medium-term value of the level of novelty.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/708/1*sspxb7Me0oXsEIwSfEnF7g.gif" /><figcaption>A criterion for the novelty of combinations of receptor activation</figcaption></figure><p>In the biological analogue, there may be some chemical signal that the neuron has to start the program of changing the forces of its synapses under the influence of external factors.</p><p>In order to understand what is the emotion of novelty, I want to present an example of an experiment from the book Development in Infancy by T.G.R. Bower:</p><blockquote>The baby can turn on the light, turning his head to the left. Most 2–3-month-old babies will turn on the light several times in a short time. Then the frequency of turning to the left falls below the level that serves as the criterion of learning. It will not increase as long as there is a stable connection between the lighting of the light and the turning of the head to the left. Let us assume that the experimenter changes the direction of communication: the light turns on when the head is turned to the right. Sooner or later the baby will turn his head to the left and the light will not light. In this case, a large number of turns to the left followed, then detection of the right circuit, which is indicated by a short high turn frequency to the right, which then decreases. If the situation does not change, this frequency will also remain low.</blockquote><blockquote>If the reinforcement scheme changes again, so that to turn on the light, first turn the head to the left and then turn to the right, then the frequency of the right turns will increase after the first attempt, when the light does not light; the frequency of the left turns will also increase, and finally the baby will turn his head to the left, then to the right and turn on the light. After a brief increase in the frequency of turns from left to right, the frequency will decrease and remain constant until the situation changes. Changing the reinforcement program causes a flash of activity, which stops after finding the right combination of turns. Thus, the child learns a fairly complex sequence of movements: for example, to the right — to the right — to the left — to the left. Growth of activity is observed every time with a change of situation. Studies show that this activity is not accidental. The baby sort of sorts through a series of hypotheses, testing sequences of movements to find out which of them is effective at the moment. Having found the correct one, he checks it several times and then stops the expressed activity. It again increases only when the previous sequence ceases to be applicable. From the behavior of babies it is clear enough that a light stimulus is not a motivating factor.</blockquote><blockquote>When testing of hypotheses begins, the baby after the first success hardly looks at the light source. He casts a fleeting glance at it to see if it is on lit or not. After making sure that the light has come on, he can show signs of pleasure, but he does not pay attention to the light source, which obviously is not the reason for his satisfaction.</blockquote><p>Let’s analyze the internal mechanisms of the child’s behavior from the example described in the example with the model:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*7Hhsrubj3jayweYiVjxmQQ.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*d4Bvq3vBazbZ2TW36aU9Pg.jpeg" /></figure><p>Take two areas, A and B, the first is responsible for the emotion of novelty, and the second is the area responsible for the assessment of the level of informational novelty. Region B has four input signals (a), and let’s say that signal Q is some kind of internal or external motive of turning the head to the left; this motivating signal activates the corresponding reflex, leading to action “1” (b), which will be the act of turning the head left. Similarly for stimulus W, but instead for turning the head to the right. Irritant E can be an internal motive for turning the head first left then right. Images of whole combinations of movements are formed in the motor cortex, which is within the cerebral cortex. So the presence of this representation of the reflex is entirely logical. And the stimulus R is the result of the appearance of a visual image of the inclusion of light. Reflexes associated with the motor system have representations both in the region B (с) and in the region A (e).</p><p>Combinations of stimuli will form new reflex arcs, new ways of spreading excitation, which will indicate a high level of information novelty. And the repetition of the same combinations with each time will show an ever lower level of novelty.</p><p>The value of the novelty level on gif can be seen on the right-hand side above the area B.</p><p>In the absence of information with a level of novelty higher than 70% in area B within 10 seconds, there is information starvation, and thus a need for novelty (f). The emotion of novelty in relation to other needs is very low, but the need for its satisfaction very quickly arises; the need for novelty is inferior in cases of necessity for self-preservation, emotions of fear are always stronger, but the need for novelty is almost insatiable. For example, this emotion controls the saccades of our eyes, at the time of viewing the images, as soon as the information transmitted from the eye’s receptors will have a low level of novelty, the eyes move. The speed with which the eye moves through the image indicates the speed of hunger-saturation cycles for the emotion of novelty. Of course, the level of novelty for different areas of the brain can be assessed in different ways.</p><p>A high level of novelty (&gt; 70%) in region B leads to the activation of the neural network (h), which shortens the plasticity in region A for a short time to form reflex arcs that will be associated with the actions that lead to satiation. Further, the activity of the “requirement cells” is suppressed (i).</p><p>Actions that lead to a satiation of the need for novelty will be repeated, even if they do not bring satisfaction, but in this case it is possible to suppress the action by the influence of other needs. For example, the need for energy conservation, or the emergence of habituation. Combinations, such as turning the head and turning the light on or turning the head and not switching on the light, will be evaluated equally for the level of novelty when they change.</p><p>Children often ask questions like “Why?” simply because the center of their emotion of novelty is associated with the performance of this action. Their experience tells them that by asking this question it is possible to receive new information and, accordingly, to meet the need for novelty.</p><p>The need for novelty is very important for learning to master the articulatory apparatus and motor skills. These systems have feedback; the actions of motor activity are immediately reflected in the receptors of the organs of vision, touch, and vestibular apparatus. And the action of the articulatory apparatus is reflected in the receptors of the organ of hearing. At the first stages of the development of animals, the combination of action and the perception of the results of this action form new reflex arcs, which, given the emotional mechanism of the need for novelty, leads to the desire to repeat these actions, as well as combinations of similar actions.</p><p>An example is children’s babble, which begins actively at about the age of six months of the baby. This stage of child development is initiated by the activation of unconditioned reflexes provoking pronouncing of sounds, but is supported by the emotion of novelty, causing the child to pronounce the same sounds or play with them.</p><p>Striving for active games, interaction with objects, etc. allows an animal and a person to not only satisfy the need for novelty, but also teaches them to manage their body and form its image.</p><h4>Love and affection</h4><p>The mechanisms for satisfying needs are universal, it is enough to single out the internal criteria of hunger and satiation. For example, consider the emotion of attachment. There are different kinds of attachments: mutual affection of mother and child, attachment to spouse, relative, etc. This emotion is expressed in the desire to be near the object of attachment or to have it in sight. In the event of loss of the attachment object, an anxious or stressful state may arise, and if the attachment object returns, then joy and other signs of pleasure are manifested. Studies and experiments of the Australian ethologist Konrad Lorenz demonstrate the mechanisms of emotion of attachment. For example, investigating the behavior of ducklings, Lorentz established that the first thing a newly hatched chick does is search for the “image of the mother”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*rSfPIaA7xr-n_8RM-X5Lhw.jpeg" /></figure><p>The only thing that a candidate had to do for the role of the mother is to move. If a moving object appears before the newly hatched ducklings, they begin to follow it. The process of writing an attachment object image into memory is called imprinting. After birth, one day with the mother is enough for her imprint to be completed. Thereafter they will completely ignore even their real mother, and they will follow the “surrogate mother” more readily.</p><p>At certain key moments of vital activity of the organism, the cells of active images are labeled and presumably this occurs with the help of the oxytocin mediator. Oxytocin is called the hormone of trust and attachment, it stands out in the dynamics of successful social contacts and interactions, breastfeeding, mutual grooming and sex. The nervous system remembers those neurons that are active at the time of increased levels of oxytocin.</p><p>The criterion of hunger for attachment emotion can be a low level of activity of “tagged” neurons, and the satiation criterion is the activity of these neurons. According to the described scheme of satisfying the needs of the nervous system, those actions that lead to a state of satiation of the need will be remembered and executed. The desire to admire the face of a loved one, a child crying at parting with the mother, the manifestation of maternal tenderness and care, grief for the loss of a kinship or loss of a pet are all the result of the work of the brain centers responsible for the emotions of attachment.</p><p>The presented model of the emotional mechanism allows us to speak about qualitative characteristics of separate emotions. Changing the speed or measure of satiation, the speed of the onset of hunger, and the overall maximum level of activity of “requirement cells” in relation to other emotional centers form the dynamics of the modeled nervous system.</p><p>The level of curiosity, the power of attachments, courage, greed, diligence, laziness, and cheerfulness, these and many other qualities of character are conditioned by the initial settings of the centers of satisfaction of needs.</p><p>In the future, developers of artificial intelligence programs will be not only focus on technical tasks, but also on tasks related to the mental qualities of the program, as is well illustrated in the film “Eva” (2011).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6VjRDWxA1l7mlSeT4CBgsA.jpeg" /></figure><p>Understanding the nature of emotions will help “program” artificial minds to execute the three laws of robotics with fidelity, it will not just be a set of instructions; the implementation of these laws themselves will bring real joy and happiness to such artificial minds.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2e01b2dc2b2e" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/simulation-of-emotions-2e01b2dc2b2e">Simulation of emotions</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Memory, memory consolidation and the “grandmother cell”]]></title>
            <link>https://medium.datadriveninvestor.com/memory-memory-consolidation-and-the-grandmother-cell-928b4e4b2d20?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/928b4e4b2d20</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[memories]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Fri, 24 Aug 2018 06:13:45 GMT</pubDate>
            <atom:updated>2018-08-24T06:15:33.663Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*zLaS2ZnlQpq32wLojmDN6A.gif" /></figure><p>The first question that arises before researchers of the nervous system and cognitive phenomena — what is memory? What is memory in the biological aspect? How does memory appear at the level of an individual neuron? And in what form is information stored in the nervous system? And now we will answer these questions.</p><h4>What is memory in the biological aspect and at the level of an individual neuron?</h4><p>There has long been an opinion in neuroscience that memory is associated with changes in the states and strength of synapses, as well as the sensitivity of the neuron itself. I described in my previous articles how these conditions change, I described habituation and sensitization using the example of the clam Aplysia and also the mechanisms of formation of conditioned reflexes. This representation of memory gave rise to the notion of connectome, which is a complete description of all connections in the nervous system, taking into account the quality and properties of these links. You can say that you, your personality is your connectome. If once you have the opportunity to completely replicate your connectome, you will get an person identical to you with the same memories and skills.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fembed.ted.com%2Ftalks%2Fsebastian_seung&amp;url=https%3A%2F%2Fwww.ted.com%2Ftalks%2Fsebastian_seung&amp;image=https%3A%2F%2Fpi.tedcdn.com%2Fr%2Fpe.tedcdn.com%2Fimages%2Fted%2F202580_800x600.jpg%3Fh%3D316%26w%3D560&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=ted" width="560" height="316" frameborder="0" scrolling="no"><a href="https://medium.com/media/65828873f8d8e48060d4fd22cef67f05/href">https://medium.com/media/65828873f8d8e48060d4fd22cef67f05/href</a></iframe><h4>In what form is information stored in the nervous system?</h4><p>If we talk about the nervous system of the mollusk, then everything is limited by the settings of the synapses in the neuron chains of the already existing reflexes. But if you turn to the nervous system of mammals, then it has a tool that allows you to create reflexes that were not found in the animal from birth. Reflexes in the nervous system are represented by the form of reflex arcs or chains of neurons, through which neural excitation is sequentially transmitted. It can be imagined that the chains of neurons form whole schemes of paths along which excitation is propagated, some paths are very large, others are smaller, and all this is subject to some restructuring and change.</p><p>When we talk about reflex arcs, we imagine a successive chain of single neurons, but in fact for the central nervous system it is peculiar that a lot of neurons are involved in the reflex arc and these transmitting chains consist of a whole array of cells. Although, in fact, each cell from this array contains practically the same information and, removing not one, but entire regions of cells, it is possible to preserve the reflex to which they belonged. This may seem irrational from the standpoint of saving space and resources, but from the point of view of the reliability of the preservation of information, this is justified. A nerve cell can be prone to death due to various unfavorable conditions, hypoxia, lack of nutrients or toxic poisoning, and it is not advisable that the death of one cell leads to the possible loss of important information for the organism. It turns out that it is impossible to estimate reliably the volumes of information operated by the brain in terms of the number of neurons and the connections between them.</p><h4>What is the nature of different degrees of depth of memorization?</h4><p>Physiologists have long been allocated two types of memory: short-term memory and long-term memory. That is, in some situations we store information for a certain short period of time and if we do not return to it, we forget it. And in other situations we remember the information for an unlimited period, if this information is important for us, then it can be preserved even immediately. Evolutionarily, these two types of memory are justified by the fact that the body does not need to store everything that happens to it, especially if this information is not important for survival and success and if this information is never repeated again, because the storage of information requires the expenditure of resources.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/1*RmOX2YeT0Yf2zYFoaJENOA.jpeg" /></figure><p>The transfer of information from short-term memory to long-term memory is called memory consolidation. On this subject, there are many studies, and in working on the model, I tried to justify the presence and imitate the depth of memorization. And by studying the materials of these studies, I came to the conclusion that perhaps we are dealing not with one kind of consolidation of memory, but with at least two types of consolidation.</p><p>The first kind is consolidation at the physiological level, these are experiments in which, for example, the chicks teach some reflex and if nothing happens, then the reflex will remain in memory, but if within twenty minutes there will be an action of electric shock or toxin, the reflex will not be stored in memory. This means that during these twenty minutes there are processes associated with storing information in memory, which can be violated, thereby preventing the preservation of information.</p><p>Changes in the structure of the cell do not occur immediately in its training, it takes some time, associated with a variety of chemical reactions and transformations, the production of the necessary proteins and the growth of new processes. But the effect of training is sometimes observed almost instantly, it is enough for us to hear some information and we already have it and can operate on it. Presumably this is due to the fact that the nerve cell has some reserve for changes, for example, in synapses, additional vesicles with a mediator can be stored, which will be allocated only in conditions of training. But after training to preserve information, it will be necessary to increase the synapse and the overall content of mediators in it, taking into account the reserve for possible training. In this case, the synapse is labeled with special proteins, and a certain complex cascade of chemical reactions is triggered, which leads to an increase in the synapse. Violation of the process of this complex cascade of chemical reactions during the time necessary for this and leads to a violation of memorization.</p><p>But in the computer model, there is no need to reflect the physiological consolidation of memory, because all changes with virtual synapses can be carried out instantly.</p><p>The next form of memory consolidation is an example of the fact that the information repeated several times is acquired for a long time, and information that is not emotionally colored and has been presented once or in a small amount will be forgotten after a while. I call this kind of consolidation hierarchical. Consider this type of consolidation using the example of mastering a conventional conditioned reflex using a computer model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*sMZDd0Cy0kSpwQ9Ldz_5CQ.jpeg" /><figcaption>Illustration from the works of Academician Ivan Pavlov</figcaption></figure><p>In the previous article, an example of how a conditioned reflex is formed on the basis of model neurons was demonstrated. Now let’s see how the conditioned reflex is formed at the level of the central system. Usually unconditioned reflexes have representations at once in several areas of the nervous system, first this representation in the region of the spinal cord, if we move further along the hierarchy, then representations in some ganglia, then in the ancient brain, for example, in the hippocampus, and finally, representation in the cortex.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xEF6h2PU8fk4-VVggL_nbg.jpeg" /><figcaption>Screenshot of the window of the simulation of the nervous system</figcaption></figure><p>On the model, you see two selected areas, they are isolated, that is, the neurons belonging to these regions do not act on each other, unless of course connected by synapses directly. I conditionally designated these two areas as “hippocampus” (the area below) and “cortex” (the area on top). There is also a reflex with the heading “R” leading to a reflex response “1”. This reflex has two representations in the hippocampus and cortex, these are two groups of white neurons. And also there are two receptor neurons associated with the “E” receptor in each region, these neurons are indifferent, and activation of them will not lead to any answer, but will create little activity. The main thing is that the neuroplasticity for these two area is different, in the area designated as the hippocampus the plasticity is higher than in the area of ​​the cortex.</p><p>If we activate the reflex with the heading “R” and the unconditioned stimulus “E”, approximately in one time interval, then a new reflex arc will occur, but most likely this will occur in the region with high neuroplasticity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*zCvAxrUyFEBYBTOYoviJbw.gif" /></figure><p>The reflex formed only in the region with high neuroplasticity will be fully functional, but can be lost because under the influence of other stimuli neurons can retrain. The hippocampus is considerably smaller than the cortex, but it contains practically all the representations that are present in the cortex. The need for the formation of new connections all the time makes use of cells that have already participated in the formation of reflexes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*nV3Q5wozCt6SjwK9QGMFnQ.gif" /><figcaption>Loss of reflex in the area of high neuroplasticity</figcaption></figure><p>If we continue training by combining the stimuli “E” and “R”, then after a certain number of repetitions a reflex arc will form in the region with a low neuroplasticity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*H2E4XcZj177u2TLypHYVmw.gif" /></figure><p>Now the reflex is more securely protected, even if it disappears from an area with high neuroplasticity, it will still be performed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*sCSSslCq2I_IhK_XsqMYmw.gif" /></figure><p>Of course, retraining of neurons in an area with low neuroplasticity is possible, but this will take more time and effort.</p><p>Thus, there are two stages of memorization: before the formation of the reflex in the cortex and after. Preservation of a copy of a reflex not only in the hippocampus but also in the cerebral cortex is a hierarchical consolidation of memory.<br>But how is it with information that is remembered immediately and for a long time, usually such information is accompanied by some emotional experiences.</p><p>Our pain receptors are associated with the brain department, called the amygdala. The amygdala controls the area of the nervous system the locus coeruleus, which consists of norepinephrine neurons. The axons of these neurons have endings in all areas of the brain, their task is to deliver norepinephrine as needed to as many cells as possible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*d5_GAgOybaDNQRedAsBRhw.png" /></figure><p>Norepinephrine in our case is a signal to increase neuroplasticity. This is a kind of “now print” team and under the influence of norepinephrine nerve cells try to make changes as soon as possible. And accordingly, when the actions of norepinephrine cease cells return to their working state, retaining all the changes that have occurred to them.</p><p>So, let’s look at this in the model.</p><p>The areas in the program not only logically separate and isolate the neurons, but it is also possible to configure some interaction scripts for these areas.</p><p>Objects of a network of a higher level are different areas of the brain. The nature of their interaction in the system is special, the activity of one region can lead to the inhibition of all neurons, or to have a modulating character, or to influence neuroplasticity in another area. Feeling of fear (tonsil activity) leads to an increase in the sensitivity of neurons in the motor cortex, modulation takes place to reduce the activation threshold of neurons. Therefore, when fear embraces us, we can run away from danger faster, our muscles do not become stronger, we just need less inner motivation for action.</p><p>In this case, there is a scenario whereby when an activity occurs in the area called “amygdala” in the “cortex” area, the neuroplasticity increases for two seconds.</p><p>Without the activation of the amygdala:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*zI_sEJn0U20eh3hzZIuYrQ.gif" /></figure><p>With the activation of the amygdala:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*ijW5LaLF9_hTCiOHgLs92A.gif" /></figure><p>Thus, it turns out that in a state of stress, learning occurs in the cortex as quickly as in regions with high plasticity. <a href="https://www.nature.com/articles/nature19325">For example, it can also be verified by studying experimental work on mice.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/838/1*aJ4NrLngxLBeP51syzshDw.jpeg" /><figcaption>P — neuroplasticity (0 ≥ P ≥ 1)</figcaption></figure><p>The hierarchical memory system can be represented as follows. There are several isolated areas, in each subsequent neuroplasticity will be smaller, and therefore, each next area will be less affected by changes. In each such area there will be a representation of the stimulus. And it will also be possible to control the memory speed by means of signals that lead to a short-term change in neuroplasticity.</p><p>On the one hand, we have a certain information filter that allows you to save information that is more often repeated, on the other hand, the ability to instantly remember important information for the organism.</p><h4>Images</h4><p>To study human memory only on the basis of the simplest conditioned reflexes is incorrect, for many it may seem very primitive with respect to the information that a person can operate on. And so we’ll talk about such a concept as an image, and how images are formed on the level of nerve cells.</p><p>The system of formation of images in the brain is associated with the phenomenon of neuronal specialization observed in biology. When studying the brain it was found that some neurons selectively react to a certain type of stimulus, that is, a certain group of neurons in your brain will be activated when you see or think about an animal, there are also groups that will react only, for example, on the face of your grandmother. There is even a name for this phenomenon of neuron specialization “<a href="https://en.wikipedia.org/wiki/Grandmother_cell">grandmother cell</a>”.</p><p>Let’s consider the mechanism of obtaining specialization by neurons on the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*TDwLgs_5bKKxshqbhoeeSw.gif" /></figure><p>We have a receptor field of 12 receptors (Q, W, E, R … V), all receptors are the same, are in equal conditions. And each receptor has a representation (a receptor neuron) in a group of interconnected cells located in a plane, similar to how it is organized in a fragment of the cortex that processes signals from the senses. This will be the area of primary processing and therefore the neurons in this area have very high neuroplasticity (P = 1).</p><p>If, for example, three receptors (Z, X, R) are activated from the receptor field, then the same principles of attraction of excitation that underlie the formation of conditioned reflexes lead to the fact that the excitation will converge in one place. Moreover, the level of the activating effect on the neuron at this site will be higher because the activating signals come from different sides almost simultaneously. Thus, the specialization of neurons is automatically formed, in this case neurons that react to the complex stimulus “Z + X + R” are determined.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*AZfb31rKKcQnZHoiumBxtg.jpeg" /></figure><p>And of course you can see that this system for forming image regions is very inaccurate. The regions of the images may intersect, and it is also possible that quite different groups of stimuli can be attached to the same region. And such an inaccuracy in the work of the brain can have a rather positive character. After all, in the recognition of images, accuracy can only interfere, and it also reveals the potential for creativity because combining previously incompatible or seeing something different in the image is possible only because these images are not so unambiguous and do not have clear boundaries. This also explains why human perception can be deceptive.</p><p>But in order to identify these areas of an increased level of activating action, it is necessary to add one more group of neurons that will be activated only in the case of obtaining such an increased effect. About the structure of the cortex and the organization of the cortex, we will talk in more detail in future articles, as well as on how to improve the image formation mechanism qualitatively.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*lKdfF3Dchgr7W0kD98CHWg.gif" /></figure><p>Now let’s connect the hierarchical system of memory and the system of formation of mental images.<br>Excitation from one stimulus can go in different ways before reaching the next level of processing. So the path that chooses nervous excitement is determined by the influence of other stimuli. This path determines which representation of a higher level will be excited and so higher in the hierarchy. Given the difference in plasticity at the levels, the images at higher levels will be more stable, namely, if the complex stimulus is not fully or noisily activated, then the area that will be more typical for the given complex stimulus will still come to activity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/775/1*0NdlcnyRCBTHg2VPJO_4iQ.jpeg" /></figure><p>Such a memory structure is similar to a tree, for example, at the first levels we can isolate neurons that will react to all faces of people, higher in hierarchy we find neurons reacting to faces of familiar people or those individuals that we have seen repeatedly and the last levels will reveal the faces of relatives , whose faces are very familiar to us. The face of our grandmother, which we can identify even in a very distorted, incomplete or noisy form, because at high levels the neuroplasticity is very low and the fuzzy selection of the “branch” will still lead to its full activation.</p><p>This memory structure explains the reason for the high speed of information extraction, the excitation from the receptors simply passes through all layers of processing, influencing each other, these signals determine the path through which excitations come to the right groups of neurons. Here there is no search of information and there are no comparison operations, with any reference information, etc. Evolution was along the path of speed, not accuracy, although the accuracy of perception was achieved by increasing the receptor fields and special ways of organizing the brain.</p><h4>LTP</h4><p>Discourse on memory would not be complete unless we touch upon the topic of long-term potency. This effect is due to the fact that when the nerve cell is exposed to a strong activating action, its sensitivity is increased for some time, and the time of this change in the cell can last from several minutes to several weeks. Advantageously, this effect is manifested in large pyramidal cells of the hippocampus, although it may be observed in other areas but in a lower concentration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/401/1*w8ZZguhDNjjt52lM9b03dQ.gif" /></figure><p>Under the influence of increased impact in the cell, a specific cascade of chemical reactions is triggered, which leads to the formation of a postsynaptic membrane of additional receptors, which increases the sensitivity of the cell. Let’s imagine how this will be implemented in the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/260/1*RRAOY_bYvfCo07CFawn9OQ.jpeg" /></figure><p>Consider the summation scheme in neurons that represents a vessel filled with a mediator, from which a continuous waste of this mediator occurs. If the level of the mediator in the vessel reaches the value “A”, then the activation of the neuron takes place. If the level of the activating effect reaches the “B” value, not only activation of the neuron will occur, but in addition, the main activation threshold “A” will decrease to the level “C”, these changes will be temporary.</p><p>This feature in the work of some cells of the hippocampus goes against the principles of habituation in the nervous system. Thanks to her, the organism can remember what happened to him a few minutes ago. If you close your eyes and for a while will be in silence, you do not even need to open your eyes to recall information about where you are and what events happened earlier. The surprising thing here is that there are no stimuli that will activate the reflex arcs leading to the neurons that are responsible for this information.</p><p>In the process of activity in the hippocampus, marking of regions and neurons by means of long-term potentiation occurs, those areas that are responsible for active images are labeled. This makes it easy to return to images that have recently been activated, even through a weaker impact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*CmiT9LAujmyP-6P76OGrvQ.gif" /></figure><p>Imagine that there are groups of neurons responsible for the images of places, for example, images: “job”, “house”, “street”, etc. Then, being in the workplace, we can receive a lot of visual, auditory, tactile images and signs indicating where we are. This leads to the fact that the neurons associated with the image of the place will be marked by long-term potentiation. Now, at a certain time interval, while long-term potentiation is in effect, a small stimulus from the root image of the place is enough to make the neurons of the corresponding image activated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/997/1*rX2Nn5S5OpTAq2SHQUPW9w.jpeg" /></figure><p>The received theory of memory, in my opinion, is very simple and laconic, and it follows from the foundations laid down in the work of neurons.</p><p>In the next article, we’ll talk about emotions, and I hope to convince you that the most complex emotions in a person’s understanding will be available for electronic brains.</p><p><a href="https://medium.com/@it.belkin/simulator-of-the-nervous-system-b0b69b4f346a">A link to the simulator of nervous systems can be found in the first article.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=928b4e4b2d20" width="1" height="1" alt=""><hr><p><a href="https://medium.datadriveninvestor.com/memory-memory-consolidation-and-the-grandmother-cell-928b4e4b2d20">Memory, memory consolidation and the “grandmother cell”</a> was originally published in <a href="https://medium.datadriveninvestor.com/">DataDrivenInvestor</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Simulator of the nervous system. Electromagnetic interaction of neurons]]></title>
            <link>https://medium.com/good-audience/simulator-of-the-nervous-system-electromagnetic-interaction-of-neurons-905d559624d4?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/905d559624d4</guid>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[neurons]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Sun, 19 Aug 2018 04:40:44 GMT</pubDate>
            <atom:updated>2018-08-22T16:14:13.688Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/953/1*J__yFVT0L1Lmlkap0bImag.jpeg" /></figure><p>In the development of the simulator of the nervous system, so far we have only dealt with well-studied aspects of the functioning of the nervous system (<a href="https://medium.com/@it.belkin/simulator-of-the-nervous-system-b0b69b4f346a">see the previous part</a>). But the complexity of modeling the nervous system and the reason why the artificial intelligence has not been created so far is the lack of a complete picture of how the nerve cell works. A detailed description of many processes occurring in the nerve cell and the nervous system as a whole, but there is no clear algorithm for their work, which could be transferred to a model or computer program.</p><p>I like the analogy about creating a theory of brain work with the assembly of a puzzle from the book of Jeff Hawkins “On Intelligence.” When composing this puzzle, we lack some elements, and some elements from another puzzle, but we have a large array of data on the nervous system and the brain, which means we have an almost assembled puzzle, so we can roughly picture the whole picture, and using our imagination to identify the missing elements.</p><p>My goal is the creation of a logical model of the work of the nervous system, one can say the creation of a sketch of what is depicted on an unfinished puzzle, and it must correspond and not contradict all available elements of the puzzle and be logically finished. To fill the gaps, some theoretical basis has been created, but it can be controversial. But for the model at this stage, the main thing is that it allows you to emulate both internal and external observable phenomena occurring in the nervous system. Within the framework of the model obtained, it is possible to explain many phenomena, such as memory and memory consolidation, emotions, neuronal specialization, and much more.</p><p>There are three types of reflex activity described by Academician Ivan Pavlov. If everything is clear with the biological mechanisms of habituation and sensitization, then with the formation of conditioned reflexes everything is not so simple as it seems. The fact is that the external manifestations of this mechanism have been extensively studied and described, and there is no explanation of how this occurs at the cellular level.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/1*L732bnl1yjTTv0xjz1OLzw.jpeg" /></figure><p>For example, we know that when the activity of two nerve centers is combined with time, a reflex arc forms between them. Subsequently, when one nerve center is activated, the excitation is transferred to another nerve center. If we formally divide such a reflex arc into segments, and consider such segments as separate elements. That it is possible to say, that at formation of a reflex arc of the conditioned reflex in each segment there is a commutation of a directed character. Each segment selects a specific direction in which there will be a transmission of a nervous excitement upon its activation. Of course, it is worth noting that this direction is not clearly defined for the segment, but can be correlated in certain values. It is even possible to talk about strengthening the transfer in a certain direction and weakening in other directions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oWYFlMUGHizuxDPDoFURPg.jpeg" /></figure><p><em>An important clarification: I intentionally exaggerate some aspects and simplify some facts, for the sake of simplicity of perception of the idea.</em></p><p>When the reflex is strengthened by repetitions, one can speak about the refinement and amplification of the transmission in the direction for each segment. This concept leads to the conclusion that if we divide the entire bark into similar segments, we will observe in each a certain orientation in the direction with varying accuracy and strength. Each segment will be part of some reflex arc of a conditional or unconditioned reflex. Presumably, this orientation in the learning process can be refined or changed.</p><p>If you turn to the neural paradigm, then it does not provide for orientation in the direction. We have a membrane and dendrites that receive signals and an axon, through which the signal is passed on to other cells after space-time summation, that is, the signal is transmitted in one direction along the axon to its endings. But in doing so, we still observe the formation of directed propagation of excitation in the brain, with the formation of conditioned reflexes.</p><p>If we consider not a single cell as a functional unit of directed commutation, but a small cell region, then one can see that the cells and their processes are very tightly interlaced, and in different directions. This gives an element of directional communication with a multitude of inputs and outputs in different directions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*rJTGT99aGBjcgnJkpBHPeg.jpeg" /></figure><p>The shape of the neuron is due to evolutionary changes. The shape of the cell was formed in nervous systems, in which only the simplest functional of nervous activity was realized. When the development of life on Earth required the addition of the formation of culinary reflexes to the set of functions of the nervous system, evolution proceeded along the path of not rebuilding the cell, but increasing their number and densely interlacing their dendrites and axons.</p><p>Thus, the property of directional commutation is distributed in groups of neurons, in the change in the strength of their synapses. The model neuron is a functional unit in modeling, and therefore an analogue in biology for it is a group of neurons, for which the phenomenon of directional commutation will be expressed.</p><p>We found out that the direction of excitation propagation is important for us, but how this direction is determined for each functional element. It is known that excitation seeks to spread to another source of excitation, and a stronger and larger focus of excitement attracts weaker ones (the conclusion of Academician Ivan Pavlov). If the functional element gets excited, then somehow it must determine the direction that will subsequently form and remain in its structure.</p><p>In my work on modeling, I started from the idea of the electromagnetic interaction of nerve cells, and this idea gave answers to many puzzles about the brain, gave a theory and a model explaining many aspects of the work of the nervous system.</p><p>Direct research on this topic has been done little, this is due to the fact that labor-intensive work is required to register changes in neurons under the influence of external electric fields. For example, an experiment conducted by neurophysiologists from the California Institute of Technology (C.A. Anastassiou, R. Perin, H. Markram, C. Koch (2011) Ephartic communication in cortical neurons — Nature Neuroscience [<a href="https://www.nature.com/articles/nn.2727">Abstract</a>], [<a href="http://www.gatsby.ucl.ac.uk/~beierh/neuro_jc/Anastassiou_etal11_EphapticCoupling.pdf">PDF</a>]), showed that extracellular electric fields generated by neurons change the characteristics of the action potentials of other neurons.</p><p>Despite the fact that the neuron has many contacts with the neighbors’ cells, the radius of its action is limited in comparison with the scale of the nervous system as a whole. It becomes unclear how the neuron commutation takes place in the formation of simple conditioned reflexes, since the distances between different representations of these or those reflexes can be counted up to hundreds of millimeters.</p><p>I.P. Pavlov explains the mechanism of formation of conditioned reflexes in the following way: <em>If two foci of excitation arise in the central nervous system, the stronger one of them “attracts” excitation from the less powerful one. If such an interaction of the strong and weak foci of excitation is combined repeatedly several times, a conditioned reflex can form.</em></p><p>Transmission of excitation in the nervous system is always accompanied by a change in electromagnetic fields. It is natural to assume that the nature of the “Pavlovsky attraction” has an electromagnetic character. Of course, there are hypotheses that neurons can interact at a certain quantum level, but the nature of these interactions is not clear, the development of quantum models should be postponed until the advent of quantum computers.</p><p>If we follow the Pavlovian ideas, then each activated neuron must determine in which direction the strongest focus of excitation exists and, subsequently, transmit excitation in the right direction. The neuron can memorize this direction and use it in the future. Here the neuron is represented as a certain commutator. The network of such switches forms a reflex arc, like an electrical circuit that can be formed, strengthened, rebuilt and destroyed. Of course, the functions of the adder are stored behind the neuron, which extends the capabilities of such a self-organized system.</p><p>To test the hypothesis, I developed a model in which a neuron, like a cellular automaton, conducts its internal calculations independently of the system, only on the basis of the collected information. First, when the neuron excitation is received, its variable q (charge) begins to change at a frequency of 0.01 c, depending on the given array of numbers characterizing the law of charge change on the surface of its membrane. A total of sixteen values, after which the neuron does not react to irritation for a short period of time.<br>For demonstration, we present four variants of the law of charge change, mainly differing in the value of the negative trace potential. It is believed that the undershoot are only a consequence of the repolarization of the neuron. In my work on the models I came to the conclusion that the undershoot is of great importance for the communication of neurons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-xzhv7iePVcHn9FXXJC2EQ.jpeg" /></figure><p>Secondly, after 0.05 second after activation, the neuron determines the direction of transmission of the excitation and transfers it. To determine the direction vector, it is logical to apply Coulomb’s law, but the microcosm of cells is not so simple and no one excludes the presence of an organoid in a neuron capable of amplifying the signals of other active neurons. Therefore, in the demonstration, we present three rules for determining the direction vector:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*6EncqlHK87tEkUNNx9WJMg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3-3MU2BmA4dduk9WQhpK_g.jpeg" /></figure><p>The first rule is the embodiment of the Coulomb law, the direction vector is defined as the sum of the interaction vectors with each other active neuron. The interaction vector is the product of the neuron charge per unit vector, divided by the square of the distance between neurons. The second rule is similar, but taking into account the inverse proportionality of the distance. And the third law without taking into account the distance between neurons.<br>Then, the signal is transmitted to all neurons in the direction of a certain direction vector, taking into account the range of the neuron, and its focus.</p><p>If there are no neurons in the direction of the vector, a new neuron will be created and excitation will be transmitted to it. Dynamic creation of neurons is of technical importance here, this makes the work of the model more visible, simplifies the miscalculation of its work.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FnpSXcHq2a24%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnpSXcHq2a24&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnpSXcHq2a24%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/daf5967063faf2685d0263850c4c363d/href">https://medium.com/media/daf5967063faf2685d0263850c4c363d/href</a></iframe><p>From the observations made, it can be concluded that Coulomb’s law is extremely ineffective, the effect of neighboring neurons is considerably stronger than that of another active site of excitation. Consequently, the “Pavlovian attraction” can not be explained by the simple interaction of electrically charged particles.<br>When applying a rule with an inverse proportionality and in combination with a small negative trace potential, one can already observe the formation of a “bridge” between two centers of excitation. Such “bridges” Academician I.P. Pavlov explained the formation of conditioned reflexes.<br>The most stable formation of connections is observed when using the rule without taking into account the distance, although in nature the implementation of such a rule is difficult.</p><h4>The final version of the model</h4><p>The nervous impulse in the entire nervous system has the same shape, and by analogy with it, the model neuron has a charge characteristic of the change in the aggregate charge on the membrane surface of the functional unit. That is, a certain law of variation of some characteristic called a charge is given.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/829/1*_v5LqwzfmLW2C4IVb-u-fQ.jpeg" /></figure><p>So the law is set in the program, the scale horizontally is this time in hundredths of a second, the vertical charge is in relative units. It differs somewhat from the spike graph in that the part of the maximum is longer in time. This is due to the fact that the spike values ​​are determined at one point of the nervous tissue during the passage of excitation, and the charge curve is the charge reflection across the entire surface of the cell or a group of cells. Just for zero on the charge scale, the state of rest of the nervous tissue is accepted. It should be noted that the law of charge change also reflects the trace potential, which was previously thought to be a consequence of some oscillation or equalization of charges separated by a membrane, but for the model this charge behavior turned out to be very important.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YgweH0dQtoBaGbVs6JYYSQ.jpeg" /></figure><p>The diagram above shows the neuron scheme. Signals from synapses of direct action (X1, X2, X3 … Xn) enter the adder (1). And if the resulting amount exceeds a certain threshold (2), then the neuron will activate. When the neuron is activated, its charge will begin to change in accordance with the established law (3). Information about these changes and the location of the neuron itself will be available to the entire system. Then, at a certain time, the mechanism for determining the vector of the preferred direction of propagation of the excitation (4) is triggered. This occurs by obtaining a certain average charge position of all active neurons, characterized by a point in space. This point will be called the pattern point, because for each combination of active cells and the state of their charges at the calculated instant of time for each neuron, the position of this point will be its own. Simply put, the charges of neurons influence the determination of the direction vector of the preferred excitation propagation, the positive charge attracts the excitation, the negative charge repels the excitation.</p><p>After obtaining the vector of the preferred excitation propagation direction (4), the force of the synapses (Y1, Y2, Y3 … Yn) is calculated. Each synapse is characterized by a synapse vector, whose origin lies at the center of the neuron and the end is connected to the center of the target neuron to which the signal is transmitted. The main parameter of the synapse is its force F, the force value is limited within certain limits.</p><p>Let us imagine that the vector of preferred excitation propagation forms around itself a certain cone whose vertex is in the center of the neuron, and the plane of the base is perpendicular to this vector. If the synapse vector falls within a region bounded by this cone, then the value of the synapse force will be increased by a certain value. And accordingly, if the synapse vector is outside the cone region, then the synapse force decreases, but the force value does not go beyond the set maximum and minimum.</p><p>Here is a very important point, this is how much the synapse force will change with each activation. This is determined by the parameter neuroplasticity P.</p><p>The formula for the new value of the synapse force will be:</p><p><strong>Fnew = Fold + I × P × (Fmax — Fmin)</strong>;<br><strong>Fmin ≥ Fnew ≥ Fmax</strong>;<br>where P — neuroplasticity (0 ≥ P ≥ 1);<br>I — the parameter determining whether the synapse vector lies within the region of increase of the synapse force (I = 1) or in the region of the decrease in the synapse force (I = -1);<br>Fold — the previous value of the synapse force;<br>Fmin — the minimum value of the synapse force;<br>Fmax is the maximum value of the synapse force.</p><p>Neuroplasticity in biology characterizes how much the neuron is submissive to changes in its structure under the influence of external conditions. For different areas of the brain, its degree of plasticity is characteristic, and it can also vary depending on certain factors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*nHl67gWWvdzcYxQJIE-VzQ.gif" /></figure><p>This example allows us to understand how conditioned reflexes are formed on the basis of a model neuron. White neurons form a reflex arc of the unconditioned reflex with the heading “R” and the answer “1”. These neurons do not change the values ​​of the forces of their synapses. Blue neurons do not initially participate in any reflex acts, they fill the rest of the space of the nervous system, and they are accidentally linked together through synapses. Therefore, if we activate one such neuron associated with the “Q” receptor, then there will be some excitation center of a random nature, after a while it will disappear without creating any answer. If we activate approximately in the same time interval an unconditioned reflex with the heading “R” and activation of the receptor “Q”, a reflex arc of the conditioned reflex will be formed. And already activation of just the receptor “Q” will lead to an answer of “1”.</p><p>Для ясности и оптимизации модели использовалось динамическое создание нейронов, которое эмулирует заполненное пространство нервной системы случайно связанными элементами. Здесь некоторый рост новых нейронов или новых соединений не моделируется, все изменения происходят только в силе синапсов, просто не демонстрируя нейроны, ранее не участвующие в каком-либо рефлекторном акте.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*E0SJruXL-dI7maw1ng2Urw.gif" /></figure><p>В следующем примере показано, как возбуждения ведут себя, когда разные центры активируются в равных условиях и с абсолютной нейропластичностью (P = 1).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*0fUs1c4QnVx1C34pKvJPOA.gif" /></figure><p>Изменение направления распространения возбуждения под действием двух центров возбуждения при абсолютной нейропластичности (Р = 1):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*XvoiwSdYClpolxfWs0foLg.gif" /></figure><p>И с низкой нейропластичностью (P = 0,1):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*PUKTRhlqQuwZNR1a2MQC7w.gif" /></figure><p>Пример работы одного из первых симуляций, разработанных мной в 2015 году:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FKCB_eYkxk98%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DKCB_eYkxk98&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FKCB_eYkxk98%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9707489ba11575186ac1d35b0dda810d/href">https://medium.com/media/9707489ba11575186ac1d35b0dda810d/href</a></iframe><p>Теперь я разрабатываю симуляцию с использованием графического процессора для расчетов, что позволяет работать с большим количеством элементов.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*3D0b0WRY0NriLwgTB7JXTA.gif" /></figure><p>Формирование условного рефлекса при низкой нейропластичности:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*tsAxpcV-myXqI6rRJAx-BA.gif" /></figure><p>Образование условного рефлекса с высокой нейропластичностью:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*jAxMLAyWw-IooM-0XbCiCQ.gif" /></figure><p>Даже если здесь неправильно применяется теория электромагнитного взаимодействия нейронов, полученная модель может продемонстрировать основные принципы работы нервной системы.</p><p>В будущих статьях мы рассмотрим более прикладные вещи, как использовать все это, чтобы имитировать память, эмоции, нейронную специализацию.</p><p><a href="https://medium.com/@it.belkin/simulator-of-the-nervous-system-b0b69b4f346a">Первая часть</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2F7b0181%3Fas_embed%3Dtrue&amp;dntp=1&amp;url=https%3A%2F%2Fupscri.be%2F7b0181%2F&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/748a06e9019942f2677bba917ef55804/href">https://medium.com/media/748a06e9019942f2677bba917ef55804/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=905d559624d4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/good-audience/simulator-of-the-nervous-system-electromagnetic-interaction-of-neurons-905d559624d4">Simulator of the nervous system. Electromagnetic interaction of neurons</a> was originally published in <a href="https://medium.com/good-audience">Good Audience</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Simulator of the nervous system]]></title>
            <link>https://medium.com/@it.belkin/simulator-of-the-nervous-system-b0b69b4f346a?source=rss-f2a911e00a6b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0b69b4f346a</guid>
            <category><![CDATA[nervous-system]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[neurons]]></category>
            <category><![CDATA[brain]]></category>
            <dc:creator><![CDATA[Andrey Belkin]]></dc:creator>
            <pubDate>Sun, 12 Aug 2018 12:31:59 GMT</pubDate>
            <atom:updated>2018-08-12T12:43:07.537Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EzDl4_jRJvKYIeE5tvg_hg.jpeg" /></figure><p>Hello, I want to share my work on creating a system that allows you to model the reflex and cognitive processes that occur in the nervous system.</p><p>The system is embodied in a simple program, created on the game engine <a href="https://unity3d.com/">Unity3D</a>. It is a kind of simulator of the nervous system, thanks to which it is possible to simulate not only simple reflexes, but also to demonstrate various phenomena in the nervous system, such as habituation, sensitization and the formation of conditioned reflexes. And also it is possible to emulate temporary and long-term memory and its consolidation, emotions and emotional behavior, and as simple emotions, for example, hunger and satiety, and more complex, such as curiosity, fear or affection. I plan to talk about all this in a series of articles. The first three articles are devoted to the basics in which I will talk about the main components of the system — <em>model neurons</em>.</p><p>In the system, we can distinguish three basic types of model neurons: a simple adder, a modulated neuron, an associative neuron. Model neurons are divided in complexity, each next neuron has an additional set of properties. This separation allows for a better understanding of the operating principles of the system, and in simpler examples, use simpler neurons.</p><h3>Simple adder</h3><p>The simplest neuron in the system is a simple adder. This is an element that can have inputs and outputs. And in the system there are several types of connections, first, these are signals from the receptors, there is a block of receptor-buttons that work in real time. Secondly, model neurons are interconnected through synapses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iCIQhBgaxUn5GFbuWTDA9g.jpeg" /><figcaption><em>Model neuron -</em> Simple adder</figcaption></figure><p>Synapses are of three types: synapses of direct action, synapses of modulating action and electrical synapses. All these types of synapses have analogues in the biological nervous system.</p><p>Synapses of direct action (1) are characterized by their strength (F), which is represented in the system as a real number. The sign of this number indicates which synapse has an inhibitory or stimulating action.<br>All signals are processed in real time, and can enter the neuron independently of each other, as occurs in a biological neuron. First of all, the signals of the synapses of direct action enter the adder (2).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2LYpBsYznh2AXLVqoJkDwg.jpeg" /><figcaption>The adder as test tube</figcaption></figure><p>The adder can be represented as follows, it is similar to a test tube in which all the portions of the mediator are collected. But when the neurotransmitter is allocated to the synaptic cleft in the biological nervous system, its continuous loss occurs. The neurotransmitter in the synaptic cleft is destroyed by a special enzyme, it can be absorbed back by a synapse or go beyond the synaptic cleft. Therefore, in our test tube there is a hole, due to which there is a permanent decrease in the total mass of the neurotransmitter (3). Parameter D (damper) characterizing the rate of decrease of the sum modulus in the adder.</p><p>If the total mass of the neurotransmitter exceeds a certain threshold (4), then the neuron is activated, but if the portions of the neurotransmitter turn out to be less than the threshold, then in time its summation level on the adder may fall to zero and no activation of the neuron.</p><p>For synapses of direct action, there are two types of neurotransmitters: stimulating and inhibitory. In the system, this is determined by the sign of the strength of the synapse. And the action of synapses different in sign are mutually neutralized. The total amount can be negative, then it will increase to zero at the set speed.</p><p>When a neuron is activated, it enters an activity state in which it stops responding to external signals and, through a certain time delay, activates all its outgoing synapses. Further, like a biological neuron, a temporary delay in recovery takes place in the neuron, at which time the neuron does not respond to external stimuli.<br>Contact synapses (7) activate a neuron if it is not in the active state or in the recovery period. It can be said that the signal, from the contact synapses, activates the neuron, if it is possible at the moment, otherwise ignored.</p><p>Let’s look at how the neuroelements work with the example of the simplest reflex: the knee reflex. The knee reflex consists of only three neurons, the receptor neuron receives a signal from the tendon receptors, then passes the signal to the intercalary neuron and then the motor neuron transmits the signal to the extensor muscle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oRxNYFuHfgtODvq-4nXeUQ.jpeg" /><figcaption>Knee-jerk reflex</figcaption></figure><p>So the reflex looks like in a simulation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/713/1*vTpxRWljWHBs5mBPq8y6lQ.gif" /><figcaption>Knee-jerk reflex in simulation</figcaption></figure><p>Example with inhibitory action. The model neuron bound to the receptor “E” has an inhibitory effect:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*ftoKOYT1nT2m0OrG53MenA.gif" /></figure><p>An example of competition of reflexes. Mutual inhibition of two different reflexes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*apGBpPxf1P8z9qg_e2kpoQ.gif" /></figure><h4>Transformation of the rhythm of excitation</h4><p>Single impulses for the nervous system are a rarity. In most cases, the signals from the receptors are a series of pulses, and the intensity of these pulses can be used to indicate the extent of the effect on the receptor. The more often the pulses are reproduced, the stronger the effect on the receptor, the stronger the pressure, if it is the pressure receptor, the higher the temperature, if it is the temperature receptor, etc.</p><p>And one of the observed phenomena in the nervous system is the transformation of the rhythm of excitation — a change in the frequency of nerve impulses when passing through the nerve center.</p><p>With the help of a model neuron it is possible to increase the frequency of nerve impulses. In the case when the level of a single exposure to a neuron is significantly higher than its activation threshold and the neuron has time to work several times before the value of the adder decreases to a level lower than the threshold.</p><p>The decrease in the pulsation frequency of nerve impulses is due to an increase in the resting time of the neuron, as noted earlier, during rest the neuron does not respond to external signals.</p><p>An example of converting pulses from rare to more frequent ones. In this case, the level of single exposure to a neuron is significantly higher than its activation threshold and it has time to work several times before the value of the adder decreases to a level lower than the threshold:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*BTmfljGvR6qMyZn4dzAnLA.gif" /></figure><p>An example of a decrease in the pulse frequency. Here it is carried out due to the increase in the rest time of the neuron, as I already noted, during rest the neuron does not respond to external signals:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*GCEXtUvWuvLCFGBrEojeSw.gif" /></figure><h4>Spatial and temporal summation</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*0_9HDE0eGWg3oYChfsvXgg.jpeg" /><figcaption>Image source (2e.mindsmachine.com)</figcaption></figure><p>The presented model of a neuron is allowed to explain how temporal and spatial summation occurs in a biological neuron. The temporal summation is manifested in the fact that a number of subthreshold pulses from this synapse, if these pulses occur frequently enough, can lead to activation of the neuron. This is due to the fact that the rate of decrease in the amount of impact may be less than the rate of replenishment of this amount.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*T6UJk_2XKOhqYQ6683nTdw.gif" /></figure><p>In biological neurons, there is a phenomenon that indicates that the summation is not only temporal, but also spatial. Even if the sum of the subthreshold signals exceeds a threshold, but these signals are fed to synapses located at a sufficient distance from each other, activation may not occur. Such a phenomenon can be modeled using several neurons, for example, each dendritic neuron may have dendrites.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/452/1*ZcvUOAT8IbCW1lQqDQ9bsg.jpeg" /></figure><h3>Modulated neuron</h3><p>There are three types of reflex activity: habituation, sensitization, and the formation of conditioned reflexes. These types of activities identified academician Ivan Pavlov and if our model does not emulate this, it means that it is not a model of the nervous system, but a model of something else.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/1*L732bnl1yjTTv0xjz1OLzw.jpeg" /><figcaption>Image source: In search of memory — Eric R. Kandel</figcaption></figure><h4>Habituation</h4><p>Habituation is a phenomenon associated with the fact that after a repeated action of an indifferent stimulus, both the animal and the cell cease to react to it. For example, the background sound to which we can get used and after a while almost cease to hear it, or if you wear the ring for a long time, you may not feel its pressure on the skin, etc.</p><p>Habituation is implemented in the following way. When the neuron is activated, it enters the state of activity, in which it ceases to react to activating factors, such as a signal, from contact synapses or exceeding the threshold of the total sum on the adder. After a while, the neuron activates all the transmitting synapses it has. After signal transmission, the time passes, which is designated as the recovery time. Then the state of activity is replaced by a state of expectation, in which the neuron can react to activating factors and again enter the state of activity. For a modulated neuron after the activity phase, there is an evaluation time over which it is determined whether the neuron will be activated again.</p><p>If re-activation occurs during this period, then repeats are counted. That is, every time the activation occurs during the evaluation, the repeat count is incremented by one, but if the activation does not take place during the evaluation, the counter is reset. This is how the activation counts, if they are committed quite often.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/793/1*Vl2dGf3ohkZN4hNdcfY3xA.jpeg" /></figure><p>And if the number of repetitions is above a certain limit, then the threshold of the neuron adder increases by a certain value. Thus, the neuron increases its threshold until it stops responding to the level of exposure that activated it.</p><p>An example of habituation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*0JkBsSkzBg_G_AQnxeCQIg.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2VHuPWHFKj5Pxg-c0rIH8g.jpeg" /></figure><h4>Adaptation</h4><p>In contrast to habituatio, it is necessary to distinguish the mechanism of adaptation. Adaptation is the ability of the cell to return to the previous level of sensitivity over time. If there is no stimulus for a long time, the force of habituation decreases, and can disappear altogether.</p><p>The speed with which the recovery takes place can be different in different cases, and sometimes it can last hours or even days, and in some cases it happens very quickly.</p><p>Adaptation example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*aZH5iw_BOKTSC6PjHqWyiQ.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1A582V7gHI7zmr66PXWkDQ.jpeg" /></figure><p>The mechanism of habituation can be represented as a defense mechanism, with very frequent activation of the nervous tissue, there is a high probability of its exhaustion, death and damage. Therefore, in order to protect the sensitivity of the neuron decreases, and it rarely begins to respond to stimulation.</p><p>On the other hand, if the neuron is not activated, then it will not perform its tasks and therefore, it would be a useless energy consumption. Therefore, there is a mechanism of adaptation, which increases the sensitivity to external stimuli, which increases the probability of activation of the neuron.</p><p>Habituation and adaptation solve the problem of cyclic transmission of excitation in the nervous tissue. Often in neural tissue, neurons are combined in such a way that they form neural excitatory transmission loops, but when ring excitation transfers do not endlessly loop, these transmissions also cease with time, due to rapid habituation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*gkNRYOcpdhKwCWmywcS3yA.gif" /></figure><h4>Sensitization</h4><p>The second type of reflex activity is sensitization. Sensitization is an increase in sensitivity to the effects of irritants, even indifferent, if this was preceded by an important event for the body.</p><p>For example, for a dog, a certain sound is an indifferent stimulus, to which it previously did not react. Then, if an unpleasant stimulus comes, for example, an electric shock, the dog will be disturbed for a while, and even an indifferent sound will cause a characteristic behavior, a protective reaction.</p><p>In order to simulate sensitization, let us turn to the works of Eric Kandel, the Nobel Prize winner in the field of physiology. He described in detail the effect of modulating synapses on the example of the nervous system of the mollusc Aplysia.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/501/1*sJM5WUW43kFjHrJzEZnPhA.jpeg" /><figcaption>Eric Kandel</figcaption></figure><p>The mollusc has a defensive reaction in response to all kinds of touching this drawing of the gills. In his experiments Eric Kandel caused Aplysia to get used to the easy touch of her siphon so that the protective reflex did not work. The protective reflex is retained for strong effects on the siphon, but with a weak effect it was absent. But if the light touch of the siphon was preceded by an impact on the tail of the mollusk, then the protective reflex worked with the same force and the gills survived.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/660/1*LcJ3x7RQQHXUBGWXTwJHuQ.jpeg" /><figcaption>Illustrations from Eric Kandel book “In Search of Memory”</figcaption></figure><p>The nervous system of Aplysia consists of a relatively small number of cells that can be identified and therefore it is possible to compose patterns of reflexes. It is possible to completely isolate the reflex arc responsible for the defensive reaction when the siphon is irritated — this is the main circuit. And the neuron chain responsible for modulation is a modulating chain that is activated when exposed to the tail of a mollusk.</p><p>Eric Kandel described in detail how modulation occurs in this case, what chemicals are involved in this process, intracellular cascades of reactions. Let’s transfer this knowledge to our model.</p><p>And so in our system there is a certain type of synapses — modulating. This type of synapse has no direct activating or inhibitory action, it acts on the level of the activating threshold. The level of activation of the neuron consists of two parts of the main and modulated. The main part of the activation threshold is what will change when habituation and adaptive. The modulated part is similar to the adder, in it all the effects from the modulating synapses are added. The resulting sum gradually decreases in modulus and tends to zero. The rate of reduction of the modulating effect is much slower than the rate at which the overall effect of synapses of direct action decreases.</p><p>The strength of the modulating synapse can be different in sign, that is, it can raise the threshold, thereby reducing the sensitivity of the neuron, or lowering the threshold, increasing its sensitivity.</p><p>The threshold level by which the neuron will be activated is the sum of the main and modulating parts, this level can not be zero or below zero.</p><p>Let’s simulate the experiments of Eric Kandel with Aplysia.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*GkKNC6zwYxdBjy9aYknUsw.gif" /></figure><p>The receptors “Q”, “W”, “E” and “R” are siphon receptors, each of which is associated with a receptor neuron. Each receptor neuron is connected by direct action synapses with an insertion modulated neuron. Which in turn is associated with the motor neuron, which signals the indicator “1”. Indicator “1” will be analogous to the action of reducing the gills in Aplysia.</p><p>Initially, the network is configured so that the activity of each of the siphon receptors will lead to the action of “1”. But with prolonged “tickling”, unhurried successive activation of the siphon receptors, after a while habituation to the intercalary neuron occurs and no response takes place. But, if the receptors of the siphon are affected more strongly, activate several receptors at once or do it more quickly, then we will see that the reaction “1” will still work. In this case, the adaptation mechanism will be very long in time, and we can neglect it.</p><p>Also in our scheme there is a receptor “F”, a receptor for the tail of a mollusk, which is associated with a neuron that has a synapse of modulating effect on an intercalary neuron. This synapse lowers the level of the neuron threshold for a certain time, which makes the neuron more sensitive and during this period the activation of even one receptor of the siphon leads to the reaction “1”.</p><p>The mechanism of sensitization in Aplisia is a prototype of the emotional state of anxiety and fear in animals with a more developed nervous system. In such animals, not individual chains of neurons participate in sensitization, but whole regions in the nervous system. The area responsible for fear and anxiety is the amygdala, with its activation occurs, the production of modulating mediators (epinephrine, norepinephrine). These mediators can exert an effect on the motor cortex, increasing the sensitivity of neurons, which increases activity in the motor cortex. This means that less internal motivation is required to accomplish certain actions, which allows you to escape from danger more quickly or more aggressively.</p><h4>Switch</h4><p>Modulated neurons also allow us to imagine how the behavior of an animal changes depending on external and internal circumstances. For example, Aplysia demonstrates a very complex behavior during the mating period, her behavior changes during this period, and possible responses to the same stimuli change. That is, we can say about the existence of a “switch” in the nervous system of the mollusk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*-gks_zY-PMtoGcuDY4LTKg.gif" /><figcaption>Switch</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*5YJ_4S5ddLtiI_zxJoeyYw.jpeg" /></figure><p>The network is configured so that when the receptor “R” is activated, a reflex response “1” occurs. But if we model this neur network by activating the receptor F, then the reflex response “2” will occur at the same receptor “R”. Modulating factor for the nervous system may be the presence of a certain hormone in the body, which explains how the behavior of the animal changes, for example, during fertility.</p><p>An example of gait switching in the nervous system:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*qUZ0avMybWg7lvLxlLJQLg.gif" /></figure><p>The above mechanisms are sufficient for modeling simple nervous systems. An example of this may be my project <a href="https://medium.com/@it.belkin/opentadpole-the-first-cybernetic-animal-1dcba8a01e9d">OPENTadpole</a>, a model of the nervous system of a tadpole frog.</p><p><a href="https://drive.google.com/file/d/0B_oxPA29sHP5cTZoTEN3ZnExVkU/view?usp=sharing">Dowland for Windows</a></p><h4>To be continued…</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0b69b4f346a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>