<?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 Ben Nashman on Medium]]></title>
        <description><![CDATA[Stories by Ben Nashman on Medium]]></description>
        <link>https://medium.com/@bennashman?source=rss-73ed60541e7a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*g6JyyEHrlbTPaCR85Nc63Q.png</url>
            <title>Stories by Ben Nashman on Medium</title>
            <link>https://medium.com/@bennashman?source=rss-73ed60541e7a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 12:14:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@bennashman/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[Nuclear Physics and Fusion: The Strongest Forces in the Universe]]></title>
            <link>https://medium.com/@bennashman/nuclear-physics-and-fusion-the-strongest-forces-in-the-universe-3cdc73e26b6b?source=rss-73ed60541e7a------2</link>
            <guid isPermaLink="false">https://medium.com/p/3cdc73e26b6b</guid>
            <category><![CDATA[nuclear-energy]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[nuclear-fusion]]></category>
            <category><![CDATA[nuclear]]></category>
            <category><![CDATA[quantum-physics]]></category>
            <dc:creator><![CDATA[Ben Nashman]]></dc:creator>
            <pubDate>Fri, 15 Feb 2019 01:37:52 GMT</pubDate>
            <atom:updated>2019-02-15T01:37:52.157Z</atom:updated>
            <content:encoded><![CDATA[<p>Why are things stable?</p><p>This seems like a silly question to ask, but it’s exactly what led to the modern understanding of the atom. In high school, we’re taught to believe that an atom has a nucleus composed of protons and neutrons, and surrounded by orbiting electrons. This isn’t entirely wrong, but I stress the word <em>believe </em>for a reason.</p><p>Think about this for a second. At the centre of an atom we have a densely populated region of positive charges, and surrounding that are negatively charged particles. I think you can see what’s supposed happen:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/513/0*W9jIonQ5TzMn_yL9" /></figure><p>According to classical physics, we should see massive explosions of electrons colliding with their nuclei! But… we don’t? This is part of what sparked the large movement toward developing a quantum theory of nature. There had to be something else!</p><p>Quantum mechanics gave us an ingenious answer to this problem. If we look at the energy diagram for an electron near a nucleus, we see that its potential energy goes toward negative infinity and its kinetic energy skyrockets toward positive infinity as we move toward the nucleus. It turns out that we end up having some balance between the two where the electron is stable around 53 pm (that’s picometers, or 5.3 x 10^-11 meters).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/442/0*MFDK9pIhN5pa6rSs" /></figure><p>Then, if we consider the <a href="https://www.youtube.com/watch?v=TQKELOE9eY4">Heisenberg uncertainty principle</a>, we see that we have to view the electron as more of a distribution than a single point. So, we get the quantum picture of an atom:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/602/0*WJHwz5fdOTuVdadz" /></figure><p>This way of looking at nuclei and electrons has been extremely successful. To further explain how electrons interact with other particles and each other, <a href="https://www.youtube.com/watch?v=ATcrrzJFtBY">quantum electrodynamics</a>was developed in the mid 20th century.</p><p>But there’s another problem that we haven’t yet considered. To see it, let’s zoom in on the nucleus:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*v1iVVEBc9hoFKXgo" /></figure><p>The nucleus is composed of protons and neutrons, collectively called nucleons. Neutrons have no charge (hence, they’re <em>neutr</em>ons) and protons have a positive charge. So the nucleus is <em>only</em> positively charged.</p><p>Then shouldn’t all the protons immediately break apart from each other? How is the nucleus held together so tightly?</p><p>Again, we can’t rely on classical physics to explain this. Instead, we need quantum mechanics, and a particularly complex field theory within it — quantum chromodynamics, or QCD for short.</p><p>QCD describes the strong interaction, one of the fundamental interactions in the universe, along with gravity, electromagnetism and the weak interaction. To understand why it’s called that, the strong interaction is about 137 times stronger than the electromagnetic interaction, and around 10³⁸ times stronger than gravity. That’s a 1 with 38 zeros.</p><p>But unlike gravity and electromagnetism, the strong force can only be experienced at extremely small distances, around 5 <em>femtometers</em> (fm) or less (1 fm = 10^-15 meters). For reference, a proton is only around 1.75 fm in diameter. That’s why we don’t see it in our everyday lives.</p><p>The strong interaction manifests itself in two ways: the color force and the nuclear force. The former is responsible for keeping the nucleons intact, and the latter describes how the nucleus as a whole is held together. Remember: both are fundamentally the same thing; they just differ in what they act on.</p><p>Every fundamental interaction in the universe corresponds to a “force-carrying particle” called a gauge boson. For electromagnetism, it’s the photon. For the weak interaction, it’s the W and Z bosons. And for the strong interaction, it’s the gluon. Whenever there’s a force, there’s a particle to “carry” that force.</p><p>If we look inside a nucleon, we see that it’s composed of smaller subatomic particles called quarks. There are <a href="http://hyperphysics.phy-astr.gsu.edu/hbase/Particles/quark.html">six different types of quarks</a>, but both neutrons and protons only include two types: up and down quarks.</p><p>Quarks have some very peculiar properties. For one, they have fractional electric charges. The up quark has a charge of +2/3 and the down quark a charge of -1/3. So, if we combine two up quarks and a down quark, we find that the charge is +1. This gives us a proton! If we combine an up quark and two down quarks, the charge is 0. This is a neutron!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/868/0*KcsqVBlPcgoHchmR" /></figure><p>You’ll notice that in the above picture each quark is coloured either red, blue or green. That’s because quarks and gluons have another charge in addition to their electric charge: they have a color charge. Electric charges require two opposing charges to be neutral: positive and negative. Color charges, however, require three different colors: red, blue and green.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/0*NxG2Qbv4a6hag2mD" /></figure><p>Note that quarks aren’t <em>actually</em> coloured. This was originally used as an analogy to explain the odd charge characteristics of quarks, but the names stuck! Richard Feynman even called his colleagues who developed this confusing terminology “idiot physicists”!</p><p>The way these quarks are held together in a nucleon is by interacting with each other through gluons. Here’s how it works:</p><ol><li>A quark with a color charge emits a gluon with that same color charge. The color charge of that quark is now different.</li><li>Another quark absorbs that gluon, which then changes its color to the gluon’s color charge.</li><li>This process is continuously repeated to make sure that all three color charges are conserved so that the nucleon remains color neutral.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*VUAVQgN8V8WYCmWZ" /></figure><p>This is a very simplified version of gluon-quark interactions. For the full explanation, which includes <em>anti-color charges</em>, I’d highly recommend reading <a href="http://webhome.phy.duke.edu/~kolena/modern/hansen.html">this overview</a>. If you want a visual of the interaction, click <a href="https://en.wikipedia.org/wiki/Color_charge#/media/File:Neutron_QCD_Animation.gif">here</a>.</p><p>So, the strong interaction arises from differently colored quarks exchanging gluons with one another. In this situation, we call this the color force.</p><p>When we zoom out to the nucleus as a whole, we see that each individual nucleon is color neutral. However, some combination of quarks and gluons will occasionally “leak away” from a nucleon towards another nucleon. This corresponds to some nuclear force leaking between different nucleons, which causes them to bind together. This is why the nuclear force is often called the <em>residual </em>strong force. If you’ve ever taken chemistry, you can view this like how intermolecular forces cause molecules to be solid, liquid or gaseous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*Ka7vjmObXaXqj1HO" /></figure><p>So that’s quantum chromodynamics in a nutshell. Of course, we only touched the surface of this incredibly complex theory. If you want a deeper understanding along with some amazing visuals, check out this <a href="https://www.youtube.com/watch?v=FoR3hq5b5yE">video</a>.</p><p>How does all this play out with nuclear fusion? We’ve all heard one time or another that fusion is the “holy grail” of energy, so why has it been so difficult to conquer?</p><p>Like was mentioned above, the strong interaction is magnitudes stronger than nearly every other fundamental force. However, it only acts at extremely short distances. So, if we want two different nuclei to bind to one another, we need to get them <em>really</em>, <em>really</em> close to one another. To do this, we have to overcome the powerful electric repulsion between nuclei. Think of this like trying to stick two magnets together by their positive ends.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YfiC6Eabl2sTO5r_" /></figure><p>If we just considered these two interactions, it would be nearly impossible to get nuclei sufficiently close to one another for fusion to occur. This is where quantum mechanics comes to the rescue again!</p><p>Since nuclei are composed of subatomic particles, they exhibit quantum mechanical properties. More specifically, each nucleus is described by a nuclear wave function. Instead of treating a nucleus as a point like centre, we can instead view it as more of a smeared wave.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jkedOQ5wPveV4iNu" /></figure><p>Occasionally, two nuclear wave functions will overlap. This can occur at distances greater than what the strong nuclear force extends to. When this occurs, the nuclei can undergo a quantum tunnelling effect. Essentially, this is when a particle passes through a barrier by having its wave function extend beyond it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DbFhr7SyeKiQEBPF" /></figure><p>So, if two nuclear wave functions overlap, they can undergo a quantum tunnelling effect to pass the electric repulsion barrier between them. This would result in them suddenly experiencing the strong nuclear force, causing them to fuse together.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GcFPiYhPZNawmInu" /></figure><p>Because of this effect, the energy required to cause nuclear fusion is much lower than what we would get if we just considered the strong interaction and electromagnetism. Still, to get nuclei in the range for this to happen requires some significant amount of energy. This is why fusion reactors have to heat their plasma samples to incredibly high temperatures and pressures.</p><p>An alternative attempt of producing nuclear fusion without the high temperatures has been in getting the nuclei <em>just</em> close enough for tunnelling to occur by taking advantage of subatomic particles called muons. In muon-catalyzed fusion, the electrons in hydrogen atoms are replaced with muons, which have the same electric charge but are around 200x heavier. If two hydrogen nuclei are then bound by a muon, they’ll be brought extremely close together so that fusion might occur. Some researchers in this field even claim that this can be performed at room temperature!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Hw4ZK7uwKLfGdtXG" /></figure><p>If you want to learn more about muon-catalyzed fusion, I’d recommend watching these two videos: click <a href="https://www.youtube.com/watch?v=aDfB3gnxRhc">here</a> and <a href="https://www.youtube.com/watch?v=g091Z8xaDqU">here</a>.</p><h3>Key Takeaways:</h3><ul><li>Quantum chromodynamics describes the strong interaction, which is manifested as the color force and the nuclear force.</li><li>Nucleons are composed of differently colored quarks constantly exchanging gluons with one another.</li><li>The nucleus is held together by the strong force leaking between nucleons.</li><li>For nuclear fusion to occur, nuclei have to overcome their strong electric repulsions to where the strong interaction becomes significant.</li><li>Muon-catalyzed fusion takes advantage of quantum tunnelling effects between nuclei that are bounded by muons. This is possible at room temperature.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3cdc73e26b6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modeling Your Brain On A Computer With Computational Neuroscience]]></title>
            <link>https://medium.com/@bennashman/modeling-your-brain-on-a-computer-with-computational-neuroscience-57596c919c70?source=rss-73ed60541e7a------2</link>
            <guid isPermaLink="false">https://medium.com/p/57596c919c70</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[computational-biology]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[consciousness]]></category>
            <category><![CDATA[simulation]]></category>
            <dc:creator><![CDATA[Ben Nashman]]></dc:creator>
            <pubDate>Fri, 15 Feb 2019 01:35:17 GMT</pubDate>
            <atom:updated>2019-02-15T01:35:17.216Z</atom:updated>
            <content:encoded><![CDATA[<p>This article gives an overview of what computational neuroscience is, how we can actually model someone’s brain and why it’s so awesome. If you’re interested in a specific topic, I broke this down into five major sections:</p><ol><li>Background</li><li>Modeling Neurons</li><li>Modeling Neural Networks</li><li>Applications</li><li>Key Takeaways</li></ol><p>But! I recommend you go through this from 1 to 5 if you <em>really </em>want to understand everything.</p><h3>Background</h3><p>In nearly every scientific field, we create models to explain how things interact and what a system can look like in the future. Meteorologists record temperatures, wind patterns, and cloud formations to predict what the weather will be like over the next few weeks. Astronomers use general relativity to determine if asteroids will someday come close to Earth. We use this information to help protect ourselves from and gain a better understanding of the world around us.</p><p>But why is there such little predicting power for biological systems? Why can’t a neuroscientist accurately predict whether someone will develop a mental health condition or how their brain will react to different stimuli?</p><p>The answer is complexity… and a whole lot of it.</p><p>The brain is the most complex system we know of in the universe. It is a complicated mess of interconnected cells constantly exchanging information with one other. This leaves us having a difficult time understanding what processes are occurring inside our heads. As Stanford neurobiologist Lu Chen says, “We know very little about the brain.”</p><p>In biology class, students learn about how neurons (brain cells) are connected via their dendrites to form larger networks, and how neurotransmitters (chemical messengers between neurons) transmit electrical signals between neurons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/696/0*sQtcUqbkmsncRx3E" /></figure><p>Other cells are also involved in supporting these neurons as they function, but they aren’t as relevant to how information is exchanged.</p><p>These complicated neural networks (not to be confused with the artificial neural networks in machine learning) eventually combine to become the different anatomical areas and lobes of the brain. We often use technologies like EEG and functional MRI to understand the general processes taking place at and between these areas. It is at this level that we best understand how the brain works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/0*S7DlL8Ma-temN55t" /></figure><p>However, diseases don’t begin at a high level. Information isn’t processed by brain lobes. To understand these concepts in detail, we can’t look to large anatomical structures but instead at the individual neurons and their connections. If we want to accurately model the brain, we need a bottom-up approach.</p><p>This is the reason for computational neuroscience.</p><p>Like any other scientific field, computational neuroscience is about making models and predicting how the system will evolve over time. To do so, we need to model both the individual neurons and the neural networks they form.</p><h3>Modeling Neurons</h3><p>When a neuron receives a stimulus, sodium ions begin entering the neuron through specific protein channels, causing a rise in the overall positive charge inside the neuron. This corresponds to an increase in the voltage across the membrane of the neuron. If enough stimuli cause the voltage to rise above a certain threshold, a wave of sodium and potassium ions propogates down the axon (the long portion) of the neuron. This is known as integration, and the resulting electrical signal is known as an action potential. If we measure the voltage over time, we get the following signal shape:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/0*15Au4hKX_lLEJbbM" /></figure><p>Note: This is a complicated process. If you want a visualization and are looking to learn more click <a href="https://www.youtube.com/watch?v=OZG8M_ldA1M">here</a>.</p><p>A good analogy is the action potential is to the brain, as a bit is to a computer. The patterns of these action potentials over time is what gives rise to useful information.</p><p>So, we can think of neurons as nodes that combine electrical input signals and give an output of their own electrical signal. Using this idea, we arrive at the integrate-and-fire model. This was proposed all the way back in 1907 by neuroscientist Louis Lapicque, and remains one of the most relevant yet simple neuron models today (how crazy is that!).</p><p>In the integrate-and-fire model, most of the biological and chemical interactions are ignored and instead we only look at the changing voltage and current. If the input current causes the voltage to rise above some defined threshold, we see a spike (i.e. action potential). Anything below that threshold and we don’t see anything.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/0*R666EaCL9ZrsXtUt" /></figure><p>If we run this model over time, we get a spike train. Here, each spike in the train corresponds to an action potential.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/0*fTr3HhJJ6UO4K6Hq" /></figure><p>When comparing this to a spike train that was recorded in an actual neuron we see many similarities!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LlbniE4c8upI4pKI" /></figure><p>Generally, integrate-and-fire models are good to work with when modeling neurons for extended periods of time or many neurons interacting with each other. Since only the input current and voltage are measured, there isn’t much need for intense computational power.</p><p>We can classify integrate-and-fire as a phenomenological model, meaning it only really tells us about what neurons do — generate spike trains with certain spike patterns. A biophysically accurate model, on the other hand, can model <em>how</em> neurons work in addition to what outputs they give.</p><p>For this, we refer to the Hodgkin-Huxley model, one of the great achievements of 20th-century biophysics. Here, all the major components of a neuron are treated as electrical elements in a circuit-like structure. Like integrate-and-fire, we’re really mainly concerned with moving charges. But here we classify all the different types of charges. For example, the ion-gradients across the membrane are treated as voltages, ion pumps are current sources, the membrane is a capacitor… In other words, we take all the chemical parts of the neuron and treat them as electrical components. This gives us a circuit diagram for the neuron.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*gvIO4pInyPwGv1Pv" /></figure><p>This functions nearly identically to an actual neuron! You could literally run this on a PC board!</p><p>Since the Hodgkin-Huxley model takes into account multiple cellular parameters, we get a much more biologically accurate action potential.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/0*t9LZw71IZgTAQ7Oo" /></figure><p>We see a fast rise in voltage once the -55mV threshold is hit, and a repolarization time over several milliseconds — very similar to an actual neuron. So the Hodgkin-Huxley model is extremely accurate in modeling the behaviour of neurons. But why isn’t it always used?</p><p>In a Hodgkin-Huxley neuron, all the components are related by a set of differential equations. This means that everything is measured in terms of how it changes with respect to everything else. With as many components as there are, we need <em>a lot</em> of computation power to run this model, especially for long periods of time and for multiple interacting neurons. This is why when computational neuroscientists scale their systems they usually use integrate-and-fire-like models.</p><p>Is there a happy medium? Can we make a compromise between scalability and biological accuracy?</p><p>In the last few decades, there has been a movement toward developing more biophysically accurate yet scalable models, where additional components are usually added to the integrate-and-fire model to improve its accuracy. Of course, we could go the other way and try to simplify biophysically accurate models, but it’s generally easier to add than to remove.</p><p>These research projects have led to new models that sit at different points throughout the accuracy-scalability spectrum. For example, the <a href="http://www.rctn.org/vs265/izhikevich-nn03.pdf">Izhikevich model</a> and the <a href="https://en.wikipedia.org/wiki/Hindmarsh%E2%80%93Rose_model">Hindmarsh–Rose model</a> (I encourage you to click the links and learn more about them! If you’re interested in other neuron models that I haven’t mentioned click <a href="https://en.wikipedia.org/wiki/Biological_neuron_model">here</a>.) The general underlying theme for nearly all these models is that chemical processes are usually represented as electrical ones (with exceptions, of course!).</p><p>But what can we <em>actually do </em>with these models? What types of systems and processes can we apply these to?</p><h3>Quick Summary</h3><ul><li>Neurons transmit information by generating electrochemical signals called action potentials.</li><li>Modeling neurons is all about generating action potentials artificially by representing them as electrical systems.</li><li>The integrate-and-fire and Hodgkin-Huxley neurons are two of the most widely used neuron models, the former being very scalable and the latter being biophysically accurate.</li></ul><h3>Modeling Neural Networks</h3><p>Modeling a single neuron is complicated, as you can see already! Now just imagine modeling a system of 100 <em>billion </em>neurons<em> </em>and the interactions between them. The problem scales exponentially. The more neurons, the more connections, the more complexity. Even only a dozen neurons interacting with each other takes up significant computational power! This is why trying to model an entire brain seems unrealistic (at least today).</p><p>Instead, we can look at specific processes and localized neural networks: How does a memory form? How do neurons learn to do things with each other? How do they control things like movement, speech and even your thoughts? Instead of trying to build a house from one block of material, we can break it down into different pieces and then put it together.</p><p>A good starting point is looking at how neural networks strengthen over time. For example, I’m guessing the first time you tried to ride a bike, you weren’t very great at it. Certain neural pathways were being activated in your peripheral nervous system and the motor cortex of your brain, but (being a new activity) these pathways weren’t very “developed”. Over time, the synapses (the connections between neurons in that pathway) became strengthened, making you a better bike rider. How does this work?</p><p>Generally, whenever someone learns something new, they have to repeat that action to become good at it. This is because of a process called synaptic plasticity, or the fact that synapses can strengthen or weaken over time. This can be explained with Hebbian learning, where each synapse is assigned a “weight”. Greater weights mean a stronger connection, and vice versa. Generally, this corresponds to the postsynaptic (output) neuron firing more easily from action potentials in the presynaptic (input) neuron. As the activity of the synapse increases, the weight increases. This is where we get the saying “neurons that fire together, wire together.”</p><p>How can we model this?</p><p>First let’s create a system of three neurons, where one has connections to the other two. It looks something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/0*7PeXayc34pUY9ZZ7" /></figure><p>For simplicity, we’ll use the integrate-and-fire model for each neuron. Here the output current generated by neuron 1 causes voltage changes in neurons 2 and 3. Let’s also assign initial weights of 0.5 to each synapse:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/954/0*FQQlq2p9Q7ePPVqw" /></figure><p>Let’s now say that neuron 1 causes neuron 2 to fire more and more. In other words, their activity increases. On the other hand, neuron 3 does not fire very often. We can then say that neurons 1 and 3 have low activity. Over time, we use <a href="https://en.wikipedia.org/wiki/Hebbian_theory">Hebbian theory</a> to determine what the future weights become. Eventually, this is what we’ll see:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/942/0*8Qf5drcXdiNNilDW" /></figure><p>Neuron 1 still gives an output of the same amount of current as before. However, the voltage change in neuron 2 is much greater than that in neuron 3 for each unit of current. This means that we’ll see more action potentials in neuron 2 than in neuron 3. If this continues over a long period of time, we may even lose the entire 1–3 synapse! Note that this is <em>very</em> similar to how artificial neural networks are trained in machine learning.</p><p>This is why we become good at tasks that we practice at, and why over time we can lose skills that we don’t use often.</p><p>Now imagine we did this for a system of thousands of neurons organized as they would be in the amygdala, a brain region highly associated with depression. If we notice that over time we see hyperactivity in certain neural pathways, that could be an indicator for depression. Of course, it is more complex than this in practice, but the fundamentals remain the same. In the clinical setting, something like this could be used to help predict if a patient would develop a mental illness, or how long it would take for them to learn a skill. There are so many possible applications, not to mention improving our understanding of the brain!</p><p>What about more complicated processes? Can we model emotions, abstract thoughts, or even consciousness?</p><p>Well yes… and no.</p><p>Theoretically, we could represent every neuron in the brain as a Hodgkin-Huxley neuron, assign weights to all the synapses, include details on related chemical interactions and feed the system various types of stimuli. In that case, then surely we could model these complicated processes.</p><p>The difficulty, however, is that 1) this would require incredible amounts of computing power (quantum computing?), and 2) we don’t even understand these processes well enough. As Lu Chen says, we just don’t know that much about how the brain works. This makes it very difficult for us to model it since we wouldn’t have much of a clue for what to even look for.</p><p>Based on our understanding so far, the best approach may be to construct models and try to correlate certain activity with that in a living brain. We can do this with certain emotions and thought processes, and try to understand how models can give rise to them.</p><p>But what about something that is more than just a process we can correlate? What about the human experience? In other words, can we develop a model for consciousness?</p><p>Consciousness is more complex than any other neural phenomenon we know of. Instead of being a single thing, like feeling happy or perceiving colour, consciousness is the culmination of all these processes into a single experience. But what exactly is it at a deeper level? How does a system of interacting neurons give rise to conscious experience?</p><p>To answer these questions, we can look to one of the more recent and successful attempts at modeling consciousness: integrated information theory (IIT).</p><p>Full disclaimer: <em>no one</em> has fully described consciousness with an accurate, evidence-backed theory. Many approaches have been successful, like IIT, but the problem is far from being completely solved (<em>yet</em>).</p><p>One of the main difficulties in trying to tackle consciousness is the fact that it is an emergent property: a property which the entire system has, but which the individual elements do not have. You’ve probably heard this phrased in another way, that <em>the whole is greater than the sum of its parts</em>. So, it’s easy to see why taking a bottom-up, reductionist approach to consciousness will lead to many obstacles very quickly.</p><p>In IIT, instead of beginning at the structure of the brain and arriving at a conscious system, it posits that there are five essential properties, or <em>axioms</em>, of consciousness:</p><ul><li>Intrinsic existence — conscious experiences actually exist</li><li>Composition — each experience is composed of multiple elements (for example, you may be feeling sitting on a chair while you see a book and a mug on a wooden table)</li><li>Information — each experience is specific and unique according to the information within it</li><li>Integration — each experience is unified; you experience the entire scene instead of each element</li><li>Exclusion — conscious experiences are definite in that they include a set amount of information for a set amount of time</li></ul><p>Each of these axioms is then explained mathematically to describe a conscious system:</p><ul><li>Intrinsic existence — the system must have its own cause-effect space; in other words, it must be able to affect itself</li><li>Composition — the system must be structured</li><li>Information — the system must have a specific cause-effect structure that processes information in a particular way</li><li>Integration — the cause-effect structure must be unified</li><li>Exclusion — the cause-effect structure must be definite and exist over a set of defined elements (math jargon for the system exists with all the different components in the experience and doesn’t need any more or fewer components; you can have an experience with a certain number of components, whether it be a book, a mug and a table, or a room full of many different things).</li></ul><p>Using these mathematical descriptions, we end up with a conscious system needing to be structured, unified, able to process information from a certain number of elements according to that structure, and able to affect its future self.</p><p>Using all these different descriptions, we end up with a single number that describes how conscious that system is. This number is represented by the Greek symbol <em>phi </em>Φ.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MprPlddABrNlkce-" /></figure><p>The larger Φ is, the more conscious the system.</p><p>How amazing is that?! We can literally give a measure of consciousness using a <em>single number</em>!</p><p>Now, it turns out that if we construct a model of the brain using artificial neurons, we achieve very high values of Φ. Why is this?</p><p>One of the ways to describe Φ is how irreducible the system is. In other words, how much can we <em>not</em> take the system apart into its separate components. Externally, the brain is very structured and symmetrical. But if we model its cellular organization using artificial neurons, we find that the system is quite asymmetrical and seemingly random. Artificial neurons are connected to hundreds or thousands of others, and the orientation and organization of each one are constantly changing (Hebbian learning). It becomes very difficult to break apart the system without losing these axioms. This gives us a <em>very, very </em>high value of Φ.</p><p>Of course, there’s much more to IIT than just what’s described here. If you’re interested or just want to learn more about what consciousness is, I encourage you to explore the integrated information theory website <a href="http://integratedinformationtheory.org/">here</a>. There’s even a page where you can construct your own systems of neurons to calculate Φ yourself!</p><h3>Quick Summary</h3><ul><li>Neural networks are composed of different neurons that are interconnected by their synapses. We can choose different neuron models depending on the size and purpose of the system.</li><li>Hebbian theory describes how synapses can be strengthened or weakened over time depending on the activity of the neurons.</li><li>To model complex neural processes, we need complex models that mimic the brain regions that the processes occur in.</li><li>We can model and measure consciousness using integrated information theory by breaking down what consciousness is into different essential properties.</li></ul><h3>Is all this actually useful?</h3><p>Besides how cool and interesting it is to just model the brain and different systems of neurons, there are a lot of extremely useful applications.</p><p>For one, the opportunities for neuroscience research seem endless. With modifiable and adaptable systems on a computer, it becomes much easier to model different processes in the brain. Researchers can change what inputs they give to the system, or what the configuration of the neurons are, and make comparisons with real measurements on living brains. Imagine if we could figure out the best way that the brain can learn new information! We could start changing how subjects are taught in schools to be more efficient, or if you want to learn a concept you can quantitatively determine how best to learn it!</p><p>Or what about medical applications? Say researchers become able to model schizophrenia on an artificial system of neurons. Instead of dangerous trial and error experiments to find out which treatments work in real patients, they can test hundreds or thousands of different treatments on a computer model with no consequences! Imagine how many new treatments for new diseases we could discover. Our understanding and treatment of mental illnesses and other brain disorders would improve exponentially!</p><p>Other than neuroscience itself, there are a few other real-world applications for computational neuroscience that are very intriguing.</p><p>The entire AI industry has fundamentals in neuroscience. The idea of creating systems of neurons and training them for computational purposes is all based on Hebbian learning in the brain. Current neural networks in machine learning are organized into different layers, and the connections between neurons of different layers are given weights. But what if there was a more effective way of processing complex information? What if there were better organizations of neurons in the brain that were able to process data and identify patterns at a faster and more efficient rate than current neural networks? What if there was an unknown process in the brain that was dramatically more efficient at learning than Hebbian theory? These are all questions that can be answered with computational neuroscience, and if we ever discover a <em>yes </em>to any of these or other questions, the applications would be beyond what we can predict.</p><p>Another interesting idea to look at is superintelligent AI. That is, a general AI that has the capability to process different types of data, identify different types of patterns and generate its own responses in different situations. One very intriguing side-effect of integrated information theory is that nearly all systems have some value of Φ. So, what if we designed an AI according to the axioms of IIT? According to the theory, that system would have consciousness, perhaps even more than ourselves. These are all ideas that we’ll have to consider very soon. It seems that the more intelligent the AI we create, the more relevant computational neuroscience becomes.</p><h3>Key Takeaways:</h3><ul><li>Neurons can be artificially modeled by looking at their basic functions and representing them electrically. For scalable systems, the simple integrate-and-fire model works well. For biological accuracy, we use the Hodgkin-Huxley model.</li><li>In systems of multiple neurons, we connect them by their synapses and modify their synaptic weights over time with Hebbian theory.</li><li>For complex processes like emotions and thoughts, we need complex models and comparisons with data from living brains.</li><li>We can model consciousness using integrated information theory and determine how conscious the system is by looking at the value of Φ.</li><li>Using all these concepts, computational neuroscience can be used to help neuroscientists understand brain function, how brain disorders arise and how to treat them, how to optimize AI systems, and for many more applications.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=57596c919c70" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How To Measure A System Without Touching It: Magnetic Resonance]]></title>
            <link>https://medium.com/@bennashman/how-to-measure-a-system-without-touching-it-magnetic-resonance-7289820b2ba8?source=rss-73ed60541e7a------2</link>
            <guid isPermaLink="false">https://medium.com/p/7289820b2ba8</guid>
            <category><![CDATA[mri]]></category>
            <category><![CDATA[magnetism]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[nuclear]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Ben Nashman]]></dc:creator>
            <pubDate>Wed, 07 Mar 2018 22:08:48 GMT</pubDate>
            <atom:updated>2018-03-07T22:08:48.486Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/810/1*49jDEG5es1E4YNweA00img.jpeg" /></figure><p>Magnetic resonance has found applications in medicine, chemistry, quantum computing and many other areas. It has allowed radiologists to peer inside the human body to help diagnose diseases such as cancer and mental illness and has helped chemists determine the 3D structures of complex molecules like proteins and lipids. But what exactly is magnetic resonance? How does it work? In other words, how do magnets give us detailed information about something without actually touching it?</p><h3>Zooming In</h3><p>If we zoom into the subatomic world, we notice that particles are generally charged. For example, protons have a +1 charge, electrons have a -1 charge, muons have a -1 charge etc. Furthermore, particles have a <strong>spin quantum number</strong><em> — </em>a form of angular momentum for particles. This isn’t the spin we’re familiar with in everyday life, like spinning a basketball, but more of an abstract description used in quantum mechanics. It’s helpful to visualize particles as spinning objects, though it’s important to note that this isn’t the most accurate description. However, for the purposes of this overview, we can consider particles as actually spinning.</p><p>The spin quantum number can have both positive and negative fractioned values. For simplification, the charge on the spin tells what direction the particle is spinning in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/344/0*h5yKijRYJFgLFv46." /></figure><p>So now we have a charged particle that’s spinning. Why is that relevant? Well, according to electromagnetism (the branch of physics dealing with magnetic fields, electric fields, and light) any moving or changing electric field creates a magnetic field. In the case of our charged, spinning particle, it produces its own magnetic field and therefore has a magnetic moment! Thus, we can think of this particle as really a tiny bar magnet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/0*gkUk9m9P50hK22n2." /></figure><p>In the case of a proton which as a +1/2 spin and +1 charge, it gives rise to a magnetic moment of about 1.4 x 10^-26 J/T, an extremely tiny value. As an individual particle, this value is negligible. So, we need to consider an entire system of <strong>spins</strong> (note: in MR terminology, “spins” are synonymous with “particles”).</p><p>We can simplify the MR experiment in three steps:</p><ol><li>Polarization</li><li>Excitation</li><li>Relaxation</li></ol><h3>Polarization</h3><p>In an undisturbed sample, all the magnetic moments of the spins will be oriented in random directions so there is no net magnetic moment. However, if we apply an external, strong magnetic field, called B0, the spins will align either <strong>parallel</strong> or <strong>antiparallel </strong>with the field. In the case of the proton, the parallel direction is of a slightly lower energy than the antiparallel direction, so a slight excess of protons orient with the field. This gives rise to a small overall magnetic moment or <strong>bulk magnetization</strong> of the sample in the direction of the field. In other words, if you go inside a strong magnet, you will have your own magnetization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/0*R4Gb5TU6IHftPE6W." /></figure><p>In magnetic resonance, all the magnetic moments of the spins cancel out except for the slight excess in the parallel direction. So, all that is detected is from that slight excess. We call this process by which spins align with the magnetic field as <strong>polarization</strong>.</p><p>Once in an external field, a spin does something quite remarkable: it precesses. Remember the final shot in Inception? The wobbling of the top is similar to what spins do in a magnetic field. Rather than just staying aligned and stationary, a spin “wobbles” around the flux lines of the magnetic field.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/753/0*KgNFOVcrU06ToxB0." /></figure><p>This precession occurs at a very specific frequency that depends on the strength of the magnetic field, known as the <strong>Larmor frequency</strong>. The Larmor frequency is also dependent on the <strong>gyromagnetic ratio </strong>(a mouthful, I know) of the specific spin. For example, a proton, which has a gyromagnetic ratio of about 43 MHz/T, precesses faster than a carbon-13 nucleus, which has a ratio of 11 MHz/T, at the same field strength.</p><h3>Excitation</h3><p>Imagine pushing someone on a swing-set. If you push off-time, they don’t gain much height, no matter how hard you push. But, if each of your pushes is on time, the person can gain height very quickly. If you’re pushing at the right moment, then you could be said to be in <strong>resonance</strong> with the system (magnetic <em>resonance… </em>hmm, it’s starting to come together). We can do the same thing with spins in a magnetic field!</p><p>As a refresh, we have spins precessing in an external magnetic field and there is a bulk magnetization in the direction of the field. If we apply a perpendicular field that oscillates at the same frequency as the Larmor frequency, then we can “push over” the spins. This is essentially applying a perpendicular radiofrequency pulse, known as the B1 field.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/642/0*TrcvJq1wbR-rf9Bg." /></figure><p>This process of pushing over spins into the perpendicular XY plane is known as <strong>excitation</strong>. Depending on the strength of the B1 field and how long it is pulsed for, spins can be flipped over at any angle. Usually, MR experiments involve 90˚ B1 pulses to maximize the amount of signal in the XY plane.</p><h3>Relaxation</h3><p>Spins in the XY plane are <em>highly </em>unstable. Think of the swing-set analogy. After pushing someone up high, the swing-set is going to want to move back to equilibrium, where the swing is stationary. I find it hard to believe that someone would continue reaching the same height with no one pushing them. In the case of our spins, they will tend toward their original polarizations. When in the XY plane, though, the spins are highly energetic. So, to move back to equilibrium, they must release that energy. This is released in the form of radio that carries the same frequency as the Larmor frequency of the spin. As the spin releases this radiofrequency, it then flips back toward its original state through a process known as <strong>relaxation</strong>. There are two main forms of relaxation: <strong>T1</strong> (spin-lattice) and <strong>T2</strong> (spin-spin) relaxation.</p><p>T1 relaxation is the process by which the bulk magnetization returns to its original maximum value. Think of it as the gaining of magnetization in the z-axis. T2 relaxation is the process by which transverse magnetization decays. It can be thought of as the loss of magnetization in the XY plane.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*UhxSfdS8ru6VrIWt." /></figure><p>Since signals are only sensitive in the XY plane, these two processes of relaxation produce an exponentially decaying signal known as a <strong>free induction decay </strong>(FID).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/773/0*4T1XHLLLvi58Hd2b." /></figure><p>This is the signal that is detected in MR experiments.</p><h3>Applications</h3><p>So that’s how a magnetic resonance signal is obtained. But what can actually be done with it? In the following, two applications are considered: magnetic resonance imaging (MRI) and nuclear magnetic resonance spectroscopy (NMR spectroscopy).</p><h3>MRI</h3><p>The driving force for MR-based imaging came about in the mid-1970s with the discovery that different tissues had different T1 and T2 relaxation times. Therefore, tissues could be characterized based on the magnitude of MR signal after a given time period. In order to maximize signal in MRI, the hydrogen nuclei (protons) in water molecules are used, since they are by far the most MR sensitive.</p><p>There are two main types of MRI scans: T1 weighted and T2 weighted, corresponding to the two different types of relaxation. In both cases, a slice of tissues is excited by a radiofrequency pulse. Then, after some time delay, the signal is measured. The magnitude of the signal at that time then corresponds to how bright that area should be in the image. In T1 weighted MRI, the signal exponentially <em>increases</em> with time. In T2 weighted scans, the signal exponentially <em>decreases </em>with time. The way tissues are differentiated are by the rate at which T1 or T2 occur, as these result in different magnitudes of signal at the given time delay.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/646/0*G6IcBe62tv7sUwn8." /></figure><p>Depending on the weighting of the image, dramatically different images can be produced of the same slice of tissues!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TkiJZDgs231Exvnj." /></figure><p>Numerous other weightings exist beyond T1 and T2 and can produce images that are sensitive to only specific tissues while ignoring others. This makes MRI a powerful medical imaging tool, especially considering it uses only radio and magnetic fields with no dangerous ionizing radiation.</p><h3>NMR Spectroscopy</h3><p>NMR was originally developed as a chemical analysis tool in the late 1940s and early 1950s. In NMR spectroscopy, samples are analyzed in terms of the frequencies that make them up. Remember how the Larmor frequency was dependent on how strong the magnetic field is? It is reasonable to assume that since the external magnetic field is constant, then all Larmor frequencies in the sample will be the same. However, this is not the case. If we look at a molecule, we notice that nuclei are surrounded by clouds of electrons. Electrons, having a negative and high gyromagnetic ratio, tend to orient themselves antiparallel with respect to the applied B0 field. These electrons with their opposing magnetic moments decrease the <em>local </em>magnetic field for the nearby nuclei. This then decreases the Larmor frequencies for those nuclei. This is known as <strong>electron shielding</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/467/0*wRG0Gh8SskUW05-T." /></figure><p>For a given molecule with multiple MR active nuclei, therefore, it will have multiple Larmor frequencies. These frequencies can be analyzed to determine the structure of the molecule and the composition of the sample.</p><p>The different frequencies are added together to form the free induction decay signal that we receive in the NMR experiment. We can “decompose” this time-domain signal into all the frequencies it is composed of using a mathematical technique called a <strong>Fourier transform</strong>. This produces a frequency-domain spectrum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/0*Tu7MvboM7hC_8dNx." /></figure><p>Here, each peak in the spectrum corresponds to a specific <em>magnetic environment </em>for the MR active nucleus of choice. For example, consider the H-NMR spectrum for ethanol:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/450/0*T5o_PqLzG7R_BGrA." /></figure><h3>So what?</h3><p>Sure, MR is interesting. But what value can it actually add? Here’s the game-changer: magnetic resonance lets us analyze and interact with a system without having to affect its physical composition, structure, and properties in any way.</p><p>Imagine you just discovered a way to modify a protein to help cure Alzheimer’s disease. Sounds pretty great, right? The difficulty is that you have to know what the original structure and composition of the protein is, how best to modify it, and how to confirm that what you modified is correct. How do you do all that without directly affecting the structural integrity or molecular makeup of the protein? Well, in each step you can use NMR spectroscopy! This is actually an entire field of itself, called (predictably) <em>protein NMR spectroscopy</em>.</p><p>Or, what if you want to use the nuclei in a molecule as the basis for qubits in a quantum computer? It seems like an intuitive approach, especially considering how easy it is to scale. To do so, though, you have to ensure that the quantum properties of the system are maintained and the qubits are not physically compromised. Sounds like a good use case for NMR! Unsurprisingly, this is called <em>NMR quantum computing</em>.</p><p>One of the best and most known applications of MR is MRI. I am doubtful that I’ll be able to give justice to the technology in such a short paragraph but nonetheless consider the following. Imagine you’re having trouble with your vision. You find objects to be blurry and you begin to develop general confusion. Your doctor recommends a CT scan, but there is no sight of a tumor or cerebral edema (swelling in the brain). Since you’ve just been given a dose of over 500 chest X-rays in a single scan, your doctor is hesitant to use any more remotely dangerous measures. But, he remembers that he can use different weightings in MRI to image various tissues and measure brain function! So, he runs a few scans and after analyzing a magnetic resonance angiography image (this is used to get a picture of all the blood vessels in a region), he determines that you have a buildup of plaque in your occipital lobe. He gives you blood thinners, and you become healthy in no time.</p><p>Essentially, whenever you have a delicate system that you want to work with while not causing any changes to it, magnetic resonance is your best tool.</p><h3>Key Takeaways</h3><p>Magnetic resonance is a powerful tool that can give us useful information of a system using only radio pulses and an external magnet. This lets us understand the system without having to compromise it in any way. To recap:</p><ul><li>MR active particles are charged and have a non-zero quantum spin number. This makes them act like tiny bar magnets.</li><li>A basic MR experiment involves polarizing particles with a magnetic field, excitation using radiofrequency pulses and T1 and T2 relaxation.</li><li>MRI and NMR are some of the most famous and useful applications of magnetic resonance and continue to be used daily.</li></ul><p>We haven’t come close to what we can do with magnetic resonance. So far, it’s been used as more of an analytical tool. Recently, however, MR has found use cases in actively growing areas like quantum computing and wireless charging.</p><p><strong>With such potential to disrupt major industries, who knows how we’ll be using it down the road?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7289820b2ba8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>