<?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 Ioannis Kalfas on Medium]]></title>
        <description><![CDATA[Stories by Ioannis Kalfas on Medium]]></description>
        <link>https://medium.com/@kalfasyan?source=rss-63890052a92------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*D4QKDPb7txP0GKg8lyscHg.jpeg</url>
            <title>Stories by Ioannis Kalfas on Medium</title>
            <link>https://medium.com/@kalfasyan?source=rss-63890052a92------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 19:43:12 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kalfasyan/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[Modeling Visual Neurons with Convolutional Neural Networks]]></title>
            <link>https://medium.com/data-science/modeling-visual-neurons-with-convolutional-neural-networks-e9c01ddfdfa7?source=rss-63890052a92------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9c01ddfdfa7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[convolutional-network]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Ioannis Kalfas]]></dc:creator>
            <pubDate>Tue, 03 Jul 2018 20:18:58 GMT</pubDate>
            <atom:updated>2021-08-04T09:59:11.520Z</atom:updated>
            <content:encoded><![CDATA[<h4>(for non-technical people)</h4><p>Our brain is weird. This weird organ has been evolving for millions of years, ever-changing and adjusting to handle novel stimuli and conditions they would encounter, looking like a bag of slimy, gooey matter folded in various ways inside our head. Yet, this organ and the billions of neurons that it is composed of are our best bet for understanding what intelligence is and how it emerges. Another good bet is to create (artificial) intelligence ourselves and then explain our own creation.</p><p>In recent years, we’ve come one step closer to creating intelligent machines. At least in terms of vision, we now have systems that can accurately classify images from thousands of categories. This is thanks to <strong>convolutional neural networks</strong> (<strong>CNNs</strong>) and the efficiency of the graphical processing units used to train them. Given a sufficient amount of images for each category to train these networks, CNNs can learn to adjust their artificial neurons’ connection strength (“weights”) in such a way that, after training, they are able to classify images that they have never even seen before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3BRLw4lsANPEfGgimG3YVQ.png" /><figcaption>Fig1. How a typical CNN looks (<a href="https://en.wikipedia.org/wiki/Convolutional_neural_network#/media/File:Typical_cnn.png">source</a>)</figcaption></figure><p>Interestingly, CNNs have been loosely inspired by our brain’s visual system in order to do that. Visual information that enters our brain through our eyes, travels through neurons in brain regions and is consumed in stages of increasing complexity, starting from simple visual representations such as edges, lines, curves etc. and continues with more complex representations such as, for instance, faces or full bodies. This results in a person understanding increasingly more complex information about the observed object. In fact, this deconstruction of information in stages of increasing complexity is what CNNs also seem to replicate. The concept is not foreign to older computational approaches to vision. It is also seen in the older, biologically inspired, <a href="https://www.google.com/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=1&amp;cad=rja&amp;uact=8&amp;ved=0ahUKEwiG8aq-0YPcAhVFNOwKHWr6DpkQFggpMAA&amp;url=http%3A%2F%2Fcbcl.mit.edu%2Fpublications%2Fps%2Fnn99.pdf&amp;usg=AOvVaw0ce8VOk9BcAa_gvLeQOg1i">HMAX </a>model that tries to explain visual processing in cortex as a “hierarchy of increasingly sophisticated representations” (more on that <a href="https://www.quora.com/What-is-the-HMAX-model">here</a>, <a href="https://www.ncbi.nlm.nih.gov/pubmed/28532370">here</a> and <a href="http://www.cell.com/neuron/abstract/S0896-6273%2817%2930509-3">here</a>).</p><p>How can we create a model for neurons of visual areas of the brain using artificial neurons in CNN layers? A “model” is a mathematical function that, given the same input as a biological neuron, can produce a similar output. In order to create it, one first needs to understand and define the input.</p><p>Here we take a look at a study performed by my colleagues and me from the Neurophysiology Group of KU Leuven (<a href="http://www.eneuro.org/content/4/3/ENEURO.0113-17.2017">publication</a>). This study looked specifically at the <strong>inferotemporal cortex</strong>, known to be one of the “late” processing stages of our visual system (<a href="https://en.wikipedia.org/wiki/Two-streams_hypothesis#Ventral_stream">ventral stream</a>) and, more specifically, from a section of cells that are more responsive to images that contain a body versus images of other categories (e.g. faces). They get excited when a body is present in the image, even if it’s headless!<br> <br> Example input is images of silhouettes on a blank background (see below)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/274/1*cTbfMUHUo6PQLvGq5u5VzQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/274/1*nycZWSVk1DJQke653Vdv7A.png" /><figcaption>Fig2. Example image shown to biological neurons (without the green frame) (<a href="http://www.eneuro.org/content/4/3/ENEURO.0113-17.2017">source</a>).</figcaption></figure><p>Biological neurons respond to images like this (Fig2), and scientists can collect neuronal responses through <strong>electrophysiology </strong>experiments. In such experiments, a monkey is shown images on a monitor, and an electrode inserted in the brain records the activity or response of neurons to the presented images.<br> <br> An example neuron was presented (through its host’s eyes, i.e., a macaque monkey named Escher) with hundreds of images of black shapes with white background. Some of the shapes were random, some shapes resembled animal or human bodies, and other shapes resembled silhouettes of other categories. The study’s goal? To find out 1) if we can use artificial neurons from CNN layers to predict the actual biological response of that neuron to silhouette images and 2) to discover which layers of a CNN (early, middle or late) model these responses best.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/325/1*nvkahbDptYPez_XF58OTBQ.jpeg" /><figcaption>Fig3. Body patch neurons would be excited looking at this…</figcaption></figure><h4>Convolutional Neural Networks</h4><p>The role of CNNs is to receive an input image and return a number of probabilities for each of the classes that it is trained to recognize.</p><p>A common CNN architecture usually includes hundreds of feature maps in the early layers and that number goes down progressively in the next deeper layers. These feature maps (as seen in the example CNN architecture image) are slices of the whole layer and consist of thousands of neurons each.</p><p>The word “feature” comes up often when talking about CNNs, and it might be confusing to understand exactly what it means. By “features,” we refer to the output of our neurons; the term “activations” is also used interchangeably with features because t is easier to grasp from a neuroscientific point of view, since it refers to how much a neuron “activates” given an input stimulus. For those interested, there are wonderful resources online that do a fine job explaining how CNNs work in greater detail (e.g. <a href="http://cs231n.github.io/convolutional-networks/">Karpathy’s course</a>, <a href="http://colah.github.io/posts/2014-07-Conv-Nets-Modular/">C. Olah’s blog</a>, <a href="https://adeshpande3.github.io/A-Beginner%27s-Guide-To-Understanding-Convolutional-Neural-Networks/">Adit Deshpande’s blog</a>, <a href="https://hackernoon.com/visualizing-parts-of-convolutional-neural-networks-using-keras-and-cats-5cc01b214e59">this hackernoon article </a>etc.). Let’s see what our neurons do.</p><p>Each neuron in a feature map can only “peek” at a small region of the input. Thus, a neuron in the upper left corner of a feature map can see a small window of the upper left corner of the input image. The neighboring neuron in the same feature map right next to it can see a small window adjacent to the first one’s window. This means that a feature map can see the whole input image by “scanning” it with its neurons. So why do we need so many feature maps? And what do all these neurons actually do?<br> <br> Each neuron is assigned a mathematical operation to perform based on which layer it belongs to. This operation involves the weights (connection strengths) of the neuron and the small window it can see. <br> <br> You might be wondering how it is possible to perform any kind of mathematical operation between weights and images. It is possible, because in the digital world everything is numbers. The weights are numbers that indicate how strongly a neuron is connected to the pixels that constitute an image (or a small window of that image). The pixels are also numbers, and in a black and white image, the black pixels are represented with the number 0, and the white pixels with the number 1. In a gray image, a pixel can take a value inside a range (that we usually define to be 0–255; see <a href="https://homepages.inf.ed.ac.uk/rbf/HIPR2/value.htm">why</a>) which indicates its brightness. One end of the range is black (darkest) and the other end is white (brightest). Color images are a bit more complex since each of their pixels has three values that indicate its brightness in a color channel. We use the Red, Green and Blue channels for displaying purposes and we thus refer to color images as RGB images.</p><p>As we said earlier, each neuron is assigned a mathematical operation as its task based on the layer it belongs to. For example, a neuron in the first convolutional layer performs a convolution operation between its weights and a small window in the input image. This convolution operation is a element-wise multiplication between weights and pixel values and then the total sum of the result. The idea is that the neuron will output a large number if it detects a pattern that looks like its weights’ pattern. By training the network, these weights actually change such that they look like patterns of the input. Therefore, you can think of a neuron as a pattern detector, checking if the area that it’s looking at contains something familiar to the pattern of its weights and outputting a “score” for this familiarity. Interestingly, neurons of the same feature map are set to have the same weights, so each feature map basically scans the whole input image looking for a specific pattern. Therefore, the more feature maps we have, the more patterns we detect.</p><p>A neuron in any subsampling, or pooling, layer takes a single value of a small window in the previous layer. Thus, a neuron of a max-pooling layer simply takes the maximum value of that small window. The idea here is that even if you shift the input image to the side, or even rotate, or scale it, the neurons in the pooling layer will still detect similar patterns. That makes CNNs tolerant to these image transformations.</p><p>The convolutional and subsampling layers progressively become smaller in size by gathering information about the image and outputting a more compressed summary of the important points they have discovered. Note that the layers receive input information from layers before them in the classical CNN architecture. There are more architectures that information “jumps” past layers creating recurrent connections, but these are not covered here. <br> <br> The image content information will end up in fully connected layers, which help us with the final classification. A neuron in a fully-connected layer, is connected to all neurons of the previous layer. The fully connected layer’s role is to create a flattened representation of all the information that previous layers have gathered about the input image. It is flattened, because neurons in these layers are not structured in any way like feature maps and one can think them as a 1-dimensional array of numbers. This flattened representation consists of even fewer neurons than the previous layers, which contributes to the final goal of reaching the number of classes that we are training the CNN to recognize. Eventually, we will have a list of scores for each of the classes as output and the highest score will indicate the most probable class.</p><h4>Mathematical modeling</h4><p>The networks used in the study were “pre-trained”. This means that they have already been trained on classifying images using a very large <a href="http://www.image-net.org/">dataset</a> and then their network weights were saved and made publicly available to <a href="http://www.vlfeat.org/matconvnet/pretrained/">download</a>. Thus, to get each layer’s activations, one provides an input image to this pre-trained network and then takes the output of each layer. One can imagine the image “flowing” in the network, just like information flows in our brain, and then we can capture it at every processing stage (layer). This captured information is what we use to create our model.<br> <br> The mathematical modeling task here is measuring the relation between the biological neuron’s response and the corresponding values of artificial neurons’ activations. This task is what statisticians call a “regression analysis”, where one tries to predict a value of one variable (biological neuron’s response) given the values of other variables (artificial neurons’ activations).<br> <br> In every regression analysis, there are two numerical matrices involved: the X matrix of predictor variables and the Y matrix of the target variable(s). In the present scenario, one can consider the whole CNN’s neurons and build a huge X matrix as the predictors, but that would be very computationally intensive even for modern day’s computers and it wouldn’t give us any idea about the differences within layers. Therefore, we built a separate regression setup for each CNN layer and an example of how this looks is shown in the image below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*XgoPgYse1B9Ktrq5rvhM1w.png" /><figcaption>Fig4. X and Y matrices for our Regression problem. Note that the images are black-n-white (binary), but to pass them as input to a CNN, we replicate them twice to end up with a 3D volume (the bnw image 3 times). The trained weights are also 3D volumes and to understand how they interact with the input, check <a href="https://cs231n.github.io/assets/conv-demo/index.html">this</a> incredible visualization from Karpathy’s CS231n course (and also read the rest of the blog for more details).</figcaption></figure><p>By building these regression analysis setups, we are essentially trying to replace a monkey’s biological neuron with all artificial neurons present in a CNN layer (<strong>X</strong>). <br> <br> Ultimately, this study found that deeper CNN layers proved to be strong models for body patch neurons, which are also deeper in our brain’s visual processing stages. While the pre-trained networks are trained using natural images, this didn’t hinder the results of this study that used silhouette images. Thus, deep CNN layers could receive an image as input and produce a response very close to what an actual neuron would produce. More importantly, the best model from one of those deep layers, could explain on average up to 80% of the variability in the responses of real neurons to silhouettes.</p><p>To build an artificial way to represent the same function as neuron C65b, we needed to use an efficient mathematical way to get Y given X. We’ll cover that in the<strong> </strong>more<strong> technical part </strong>(coming soon). There, we’ll go over our modeling options and why it’s not such an easy problem that can be tackled with simple linear multiple regression.</p><h4>THE END</h4><p><em>(more technical part on mathematical modeling to follow)</em></p><blockquote><em>DISCLAIMER: I am an author of eNeuro’s publication </em><a href="http://www.eneuro.org/content/4/3/ENEURO.0113-17.2017"><em>here</em></a><em>. All data presented here (example images) belong to KU Leuven and for more details on the subject you can refer to eNeuro journal, or KU Leuven.</em></blockquote><p>Yannis</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9c01ddfdfa7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/modeling-visual-neurons-with-convolutional-neural-networks-e9c01ddfdfa7">Modeling Visual Neurons with Convolutional Neural Networks</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Connectionist approach to Deep Learning]]></title>
            <link>https://medium.com/@kalfasyan/my-connectionist-approach-to-deep-learning-part-1-a9b190356295?source=rss-63890052a92------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9b190356295</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[deep-neural-networks]]></category>
            <category><![CDATA[computational-biology]]></category>
            <dc:creator><![CDATA[Ioannis Kalfas]]></dc:creator>
            <pubDate>Wed, 12 Apr 2017 12:39:55 GMT</pubDate>
            <atom:updated>2018-04-22T10:46:39.731Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<h4>My background and the path that led me to study Deep Learning.</h4><p>There has been a lot of fuss about Convolutional Neural Networks and Deep Learning in general. Nowdays, you meet these concepts in almost any scientific field that one could possibly work on. In my case, this field is Neuroscience. The success of Deep Neural Networks has provided me the opportunity to satisfy my curiosity concerning Neural Networks and the relationship between Artificial and Biological ones.</p><p>Since I gained this interest in understanding Human Intelligence, the first logical thoughts that I processed pointed towards understanding how the Brain works. This seems to be the “root of all evil”. However, with a background in Informatics (BSc) and a specialization in Educational Software, one could presumably only take small steps to the direction of Neuroscience. It seemed reasonable to find a track that offers flexibility on choosing courses from both Biology and Computer Science. This flexibility I found in the MSc programme of Machine Learning at KTH University of Stockholm, Sweden. There I managed to follow courses that tought me how it’s possible for machines to learn on their own(Artificial Neural Networks, Machine Learning etc.) as well as courses that showed me how biological organisms learn/behave, such as Quantitative Systems Biology and Biomedicine (KTH), Neuroscience (Karolinska Institute) and Human Behavior (Stockholm University).</p><p>The most striking approach, that seemed to be the common ground for all the fields I studied, is ‘<a href="http://www.iep.utm.edu/connect/">Connectionism</a>’ .</p><blockquote>Connectionism is a set of approaches in the fields of artificial intelligence, cognitive psychology, cognitive science, neuroscience, and philosophy of mind, that models mental or behavioral phenomena as the emergent processes of interconnected networks of simple units. The term was introduced by Donald Hebb in the 1940s. There are many forms of connectionism, but the most common forms use neural network models. ~<a href="https://en.wikipedia.org/w/index.php?title=Connectionism&amp;oldid=766065332">Wikipedia</a></blockquote><p>Reading on Connectionism, one can gain a wider perspective on the emergence of Artificial Neural Networks in the past decades. Almost every Deep Learning enthusiast nowadays is a connectionist without realizing. With Deep Learning becoming very popular only after 2012 and not quite reaching many of my professors at KTH before I graduated from my MSc., the most attractive Connectionist approach for my <a href="http://kth.diva-portal.org/smash/get/diva2:868833/FULLTEXT01.pdf">Thesis</a> was Spiking Neural Networks in the field of <a href="http://www.scholarpedia.org/article/Encyclopedia:Computational_neuroscience">Computational Neuroscience</a>. It would enable me to study their inner workings based on neuroscientific knowledge, but through a computational lens. Therefore, I focused on studying Spiking Networks in the department of Computational Biology (KTH), building realistic graphs of neuronal connections from a structurally organized network (in horizontal cortical layers and vertical columns i.e. minicolumns, hypercolumns) and trying to interpret their dynamics. Each artificial neuron, and its synapses to other neurons, is modeled based on biological observations, which is also true for the interconnections of groups of neurons across structural modules. This gave me experience in using both Python and Spiking Network Simulators, like (Py)<a href="http://www.nest-simulator.org/">NEST</a>, in larger scale projects. Here is an <a href="http://www.nest-simulator.org/introduction-to-pynest/part-1-neurons-and-simple-neural-networks/">introduction</a> to satisfy your curiosity.</p><p>The emergence of Deep Learning was pretty obvious by the time of my graduation (Sep 15&#39;), not only in the vast field of Artificial Intelligence, but also in (Computational) Neuroscience and its respective job boards/mailing lists (e.g. <a href="https://mailman.srv.cs.cmu.edu/mailman/listinfo/connectionists">connectionists</a>, <a href="http://www.tnb.ua.ac.be/mailman/listinfo/comp-neuro">comp-neuro</a>). Deep Learning introduced neural networks that developed similar feature representations in a hierarchical way (Figure 1), just like our Visual System. This seemed (and later proved) to be our best guess for modeling the brain from low to mid/high level scales (single-cell recordings, multi-units and fMRI). I rapidly took advantage of an exciting opportunity to use Deep Neural Networks with biological data, in order to model neurons from a mammalian Visual Cortex (macaque body patches, IT Cortex), in the department of Neurophysiology at KU Leuven.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*jl3je_hGCESg-G8dy3Z9Yg.png" /><figcaption><a href="https://image.slidesharecdn.com/yannlecun-131025100724-phpapp02/95/yann-le-cun-6-638.jpg?cb=1382695742"><strong>Figure 1</strong>.</a> Deep Neural Networks learn hierarchical feature representations through their units’ weights. Source: Y LeCun presentation, adapted from Zeiler &amp; Fergus 2013</figcaption></figure><p>There, I use the activations of Deep Neural Network Layer units upon “seeing” the same images that were presented to the Macaques, as predictors of the recorded neurons’ spiking activity (firing rate). More technical details on my project will be provided in future posts.</p><p>Yannis</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9b190356295" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Convolutional neural networks as models of primate vision.]]></title>
            <link>https://medium.com/@kalfasyan/convolutional-neural-networks-as-biological-models-of-primate-vision-99aa36200d74?source=rss-63890052a92------2</link>
            <guid isPermaLink="false">https://medium.com/p/99aa36200d74</guid>
            <category><![CDATA[deep-neural-networks]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Ioannis Kalfas]]></dc:creator>
            <pubDate>Thu, 31 Mar 2016 10:55:24 GMT</pubDate>
            <atom:updated>2017-09-20T15:07:15.476Z</atom:updated>
            <content:encoded><![CDATA[<p>Recent advances in the field of Artificial Neural Networks (ANNs) have facilitated the establishment of very successful object recognition models in Artificial Intelligence. Deep Neural Networks, and more specifically, Convolutional Neural Networks (CNNs), are inspired by the primate brain’s visual system and they apparently rival the representational performance of the Inferior Temporal (IT) Cortex on visual object recognition tasks [1], while outperforming every other computer vision model on similar tasks [2].</p><p>The IT Cortex is crucial for visual object recognition and is considered to be the final stage in the ventral cortical visual system [3]. Functional imaging (fMRI), in humans and monkeys, and single cell studies in monkeys reveal that the IT cortex contains representations of bodies. Face- and body-selective regions were identified in human visual cortex using fMRI [4,5]. Monkey-fMRI studies [6–10] show also body-category selective patches in macaque IT.</p><p>Similarly to the primate brain’s visual hierarchy, CNNs’ information processing relies on the increasing complexity of the neural representations along the visual pathways, which is expressed in the layers of the networks (Figure 1). Thus, lower layers contain local visual feature detectors that have small receptive fields (RFs) and higher layers, units have more global RFs that receive inputs from units of the previous layer. The convolutional architecture with the in-between local pooling stages, allows for object translation, rotation and scale invariance; all of which, are important attributes of the primate brain and non-trivial problems for earlier computer vision models [12], that are now robustly outperformed by CNNs. Therefore, CNN models are seemingly promising candidates for modeling primate vision.</p><p>Already, methods exist that can visualize the RFs and selectivities of the units in the different layers of a CNN [13–17]. These methods principally can show how the response is driven through the network according to a particular stimuli. Furthermore, various studies have already assessed CNNs as biological vision models and they find that models which exploit representational spaces that are alike those of the IT cortex [18], usually perform better at object recognition [19,20].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7JlTl74HN8V6AJxR." /><figcaption><strong>Figure 1.</strong><em> Deep supervised learning of CNNs produces feature representations that match the corresponding representations found in the Primate Visual System.</em></figcaption></figure><h4>References</h4><ol><li>Cadieu, Charles F., et al. “Deep neural networks rival the representation of primate it cortex for core visual object recognition.” PLoS Comput Biol 10.12 (2014): e1003963.</li><li>Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. “Imagenet classification with deep convolutional neural networks.” Advances in neural information processing systems. 2012.</li><li>Gross, Charles G., C. E. Rocha-Miranda, and D. B. Bender. “Visual properties of neurons in inferotemporal cortex of the Macaque.” Journal of neurophysiology 35.1 (1972): 96–111.</li><li>Tsao, D.Y., and Livingstone, M.S. (2008). Mechanisms of face perception. Annu.Rev.Neurosci. 31, 411–437.</li><li>Peelen, M.V., and Downing, P.E. (2007). The neural basis of visual body perception. Nat.Rev.Neurosci. 8, 636–648.</li><li>Tsao, D.Y., Freiwald, W.A., Knutsen, T.A., Mandeville, J.B., and Tootell, R.B. (2003). Faces and objects in macaque cerebral cortex. Nat.Neurosci. 6, 989–995.</li><li>Pinsk, M.A., DeSimone, K., Moore, T., Gross, C.G., and Kastner, S. (2005). Representations of faces and body parts in macaque temporal cortex: a functional MRI study. Proceedings of the National Academy of Sciences of the United States of America 102, 6996–7001.</li><li>Bell, A.H., Hadj-Bouziane, F., Frihauf, J.B., Tootell, R.B., and Ungerleider, L.G. (2009). Object representations in the temporal cortex of monkeys and humans as revealed by functional magnetic resonance imaging. Journal of neurophysiology 101, 688–700.</li><li>Bell, A.H., Malecek, N.J., Morin, E.L., Hadj-Bouziane, F., Tootell, R.B., and Ungerleider, L.G. (2011). Relationship between functional magnetic resonance imaging-identified regions and neuronal category selectivity. J.Neurosci. 31, 12229–12240.</li><li>Popivanov, I.D., Jastorff, J., Vanduffel, W., and Vogels, R. (2012). Stimulus representations in body-selective regions of the macaque cortex assessed with event-related fMRI. Neuroimage. 63, 723–741.</li><li>Popivanov, Ivo D., et al. “Heterogeneous single-unit selectivity in an fMRI-defined body-selective patch.” The Journal of Neuroscience 34.1 (2014): 95–111.</li><li>Lowe, David G. “Object recognition from local scale-invariant features.”Computer vision, 1999. The proceedings of the seventh IEEE international conference on. Vol. 2. Ieee, 1999.</li><li>Zeiler, Matthew D., and Rob Fergus. “Visualizing and understanding convolutional networks.” Computer vision–ECCV 2014. Springer International Publishing, 2014. 818–833.</li><li>Girshick, Ross, et al. “Rich feature hierarchies for accurate object detection and semantic segmentation.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2014.</li><li>Simonyan, Karen, and Andrew Zisserman. “Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556(2014).</li><li>Tsai, Chuan-Yung, and David D. Cox. “Measuring and understanding sensory representations within deep networks using a numerical optimization framework.” arXiv preprint arXiv:1502.04972 (2015).</li><li>Zhou, Bolei, et al. “Object detectors emerge in deep scene cnns.” arXiv preprint arXiv:1412.6856 (2014).</li><li>Tanaka, Keiji. “Inferotemporal cortex and object vision.” Annual review of neuroscience 19.1 (1996): 109–139.</li><li>Yamins, Daniel L., et al. “Hierarchical modular optimization of convolutional networks achieves representations similar to macaque IT and human ventral stream.” Advances in neural information processing systems. 2013.</li><li>Khaligh-Razavi, S. M., and N. Kriegeskorte. “Object-vision models that better explain IT also categorize better, but all models fail at both.” Cosyne Abstracts, Salt Lake City USA (2013).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99aa36200d74" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>