neuroscience – Thinking with Nate

Brain-Like Computing

(This post’s featured image is Slicing the Connectome by Alfred Anwander, a visualization of the deep structure of the human brain. It shows flowing bundles of thin fibers, color coded by their direction of travel and which parts of the brain they integrate. Used under the Creative Commons ShareAlike 4.0 International license)

Artificial Neural Networks (ANNs) are a revolutionary technology. They’re the foundation of Deep Learning (DL), the technique that’s brought us AI systems that understand images, text, and speech, can beat us playing chess, go, and StarCraft, or even drive a car. These models are supposedly “brain-like.” We call them “neural” after all, and they can do many things people do, but how much are they like us, really? Some argue that in a few years we’ll scale this technology up and find that AI is smarter than we are! But if we take a closer look at just how “brain-like” these systems are, there’s good reason to doubt that. Is “bigger” all we need, or is a human mind something fundamentally different?

It’s worth starting with how brains and ANNs really are similar to each other. ANNs were inspired by observations of actual brains, both studying their anatomy under a microscope, and watching patterns of electrical activity in living subjects. Our understanding of how brains work is still very crude, but there are some general principles we can glean that seem very important. The brain is a vast network of small, relatively simple modules (neurons). Each one is connected to a few or even hundreds of other neurons in a vast tangled network. These neurons sometimes “spike” with electrical activity, and when one does it often nudges its neighbors to spike, too. Information can flow, transform, and cascade as it propagates through the network. Often many neurons fire together, and it seems that the meaning they represent is actually held in the collective firing patterns across the whole network. That suggests what matters in terms of scale isn’t how many neurons there are, but how many different combinations of neurons, which is vastly more. Learning seems to involve the neurons “re-wiring” with each other, adding or removing connections (synapses) to other neurons, or changing their strength.

Computer scientists were inspired by this picture of the brain, seeing it as a sort of programmable circuit. Each neuron does just a small, simple operation, but by wiring them together into a large network you can make more complex functions. Importantly, such a network could represent nearly any function, depending on the particular way the neurons get wired up together. For this reason, ANNs are often known as “universal function approximators.” Another common way of understanding ANNs is through the math used to implement them. Each neuron does a simple operation: it takes in signals from all the synapses that feed into it, multiplies those signals by “weights” that determine how much this neuron listens to each of its neighbors, adds them up, then performs an “activation function” on the result. Often, that’s as simple as returning the final sum, or zero if that sum is negative. ANNs use large arrays of identical neurons, and the final output is basically just a math expression that does all those additions and multiplications. The synapse weights are variables (or “parameters”) in this expression, and “training” a neural network is simply finding the right values for all the variables so that the math expression resolves to the answer you want, rather than something else.

But there are an infinite number of possible functions. If an ANN can represent any of them, how do you get it to be the function you want and not just some random garbage? To start with, you need to have a whole bunch of data. Either you have some task you want the neural network to do and correct answers you want it to learn (say, the name of what appears in an image, or a translation from one language to another), or you have many examples of something you want the network to generate more of (like text or images). Then, you can use an external process that evaluates the network, compares the results to the target, tweaks the connections between neurons to make the result a little better, and then does that over and over again. Researchers have found clever ways to make this training process as efficient and reliable as possible, but it can still be painfully slow and expensive, using vast amounts of data and computing power. Generally you never get a perfect result, so you carry on training until the outputs are “good enough” to satisfy the programmer. Then training stops. All the trainable parameters in the ANN are locked down. What you have left doesn’t learn at all, it’s just a math function that does… whatever it was configured to do.

This already suggests a massive difference with the human brain. For a human being, there is no “right answer” to most challenges we face. There may be serious consequences for getting the answer wrong, but this may not be obvious in the moment. This makes learning how to respond vastly harder, because we have to set our own goals, and intuit for ourselves if we got it right, now and across time. There’s no “external” training process to rewire the neurons, either. Each neuron manages its own connections, somehow knowing which ones to strengthen or weaken without any top-down picture of the network they are making or the problem they are trying to solve. There’s no “training phase” where the brain gets configured once and for all. No, we learn continually as we act, which means we adapt continuously, changing our understanding and our strategy in context (unlike ANNs). It’s even more complicated than that. Animals with brains seem to have an “offline” learning process, too; dreams play an important role in memory consolidation and mastering new skills. It’s not clear how that relates to other kinds of learning and it seems very different from the training process we use for ANNs.

Another major difference is how ANNs are organized. DL’s breakout success came from computer vision models inspired by the visual cortex. Brain researchers showed that animal vision depends on layers of neurons, arranged in a hierarchy. The low-level ones detect primitive patterns like edges of a particular orientation. Then higher layers use those edge detectors to find higher-level patterns, like shapes and objects. For a long time, computer vision systems worked this way, too (though, “vision transformers” are starting to change that). Much effort was put into making systems with more layers in order to learn more complex functions, which is where the “deep” in “Deep Learning” comes from. The hope was that if we made them much larger, they would perform much better, and so far that’s mostly been true.

But are deep, hierarchical ANNs really brain-like? They seem to be vastly smaller and simpler, actually. For one thing, it’s estimated that the human brain has 86 billion neurons with 150 trillion connections between them. By contrast, GPT5 is estimated to have 2-5 trillion parameters, but this is likely not a fair comparison. Each living neuron is vastly more complicated than the simple “weighted sum” of a neuron in an ANN. We don’t fully understand what neurons do (not to mention all the other kinds of cells in the brain, which may also play critical roles in thinking and learning), but they are living cells that autonomously manage their activations and their relationships with other cells. Unlike ANNs, where each neuron produces a single value and all neurons activate at the same time, biological neurons communicate in complex ways across multiple channels with patterns that change over time, at multiple frequencies. That means each biological neuron may be doing much larger and more complex tasks than their artificial counterparts. But aside from being a few orders of magnitude bigger than an ANN, the brain also has a much more complex structure.

The outermost part of the human brain is called the neocortex, and it’s where most of our “higher thought” happens. It’s broken down into regions which each serve a specialized function, like vision, language, logic, or social interaction. But these regions don’t look very different from each other, anatomically speaking. We mostly discovered this structure by measuring electrical activity while a subject sits in a scanner thinking about something. The neocortex is more or less uniform in its structure, but different parts are assigned different roles, and the deep structure of the brain connects them up in a very particular way. Human brains have a highly stereotyped layout, with roughly the same set of faculties interacting with each other in roughly the same way in every person. Except, of course, when something goes wrong. Brains are also very plastic, and can reorganize around damage, and adapt to patterns of use. For instance, people born blind often show activation in the visual cortex when they read braille. I like to think of the neocortex as a sort of general purpose “cognitive fabric.” Evolution, development, and learning subdivide it into modules, then compose those modules into a particular cognitive architecture that suits a species’ lifestyle. Brain cells are like an evolved platform for building brains and minds of all sorts.

If the “universal function approximator” metaphor applies to brains, then brains have vast numbers of them, arranged into a complex network, where information flows through paths that diverge and converge and loop back on each other in very particular ways. It would be like a whole computer, specialized for the purpose of generating the human mind, built from these programmable circuits. If so, then what makes our minds “human-like” is not the fine structure of the brain that inspired ANNs, but the coarse structure that determines the shape and flow of our thoughts. In nature, that design is evolved, but flexible, and it continuously adapts to meet demands. By comparison, a programmer working with ANNs chooses an architecture by intuition, it’s usually very simple, and then it never changes. Researchers are continually exploring different kinds of neural modules and ways of composing them together, but they’re all much simpler than actual brain networks, which really aren’t the inspiration any more. AI researchers have found ways to do useful work with ANNs, and don’t care whether they resemble brains or not.

But we shouldn’t commit to the “function approximator” metaphor too quickly. Some would argue the brain is nothing at all like a function! An ANN is a function. It is a math expression that describes how to turn inputs to outputs. When a programmer evaluates that expression, they get an answer, but otherwise the ANN does nothing. It’s just a pile of math. By contrast, the mind is always active, whether it’s observing the senses, controlling the body, thinking hard, or just daydreaming. In some sense, the brain has a narrow interface with the body and the rest of the world through the brain stem. If the brain implements a function, this could be where the inputs and outputs pass… except, it’s not clear where the mind ends and the body begins! The nervous system extends throughout the body, sensing, controlling, and interacting with the various organ systems, often without checking in with the brain at all. The human gut (and the microbes that live there) play a role in cognition and can shape our mood and behavior. It’s probably not the only organ system that does that, either. The mind is deeply integrated with the body, driven by its needs, cravings, and instincts. The mind is also creative, generating its own inputs, setting goals, identifying problems to solve and strategies to solve them. This doesn’t seem very function-like at all!

By using the same low-level principle as the brain, ANNs let us automatically design functions to solve a wide range of difficult real-world problems. Often, we do this by observing humans solve some problem, and copying what they do. In short, we choose what human skills to automate, and we make functions that simulate those skills. The AI doesn’t “figure out” the answers; we have to already know the answers, and then bake them into the AI. Using many processor cores running in parallel, we spend the equivalent of hundreds of years of compute time to train our models to approximate what a human can do effortlessly by intuition. Once we pay that up front cost, then that knowledge is baked into the network, unchanging until we decide to rebuild it. In short, all the things that make the human mind open-ended, creative, and continuously adaptive are missing from ANNs. They don’t have the high-level architecture that makes human minds so robust, capable, and well-rounded, and we don’t know how to make anything like that. We have no idea what “thought” even is, let alone how to create it.

There’s no reason to think AI couldn’t become more brain-like and mind-like over time. There’s nothing magical about our biology. But it seems absurd to think that scaling up our current designs will get us there. Modern AIs are very good at mimicking human data. They appear smart, because they reflect and reshape the intelligence we feed into them. But they simply do not do all the extraordinary things our brains do. Most importantly, they do not continuously set their own goals, improve their design, or adapt to changing needs through evolution, development, and learning. No, every innovation in AI is an act of human cleverness imposed onto the machine. This is why I think we have a long way to go. What we’ve built is a toy compared to an actual human brain. We’ve successfully borrowed a few of nature’s tricks, but there’s vastly more there, and we hardly even know what we’re missing. We will need many more clever insights, and those do not come often or easily.

Do Brains do Back Propagation?

I’ve been thinking about this article Ulrike Hahn shared with me recently (alternative source here). Apparently, I have strong opinions about why we shouldn’t say that the brain is doing something “backprop-like” when we learn!

Before we start, the key thing to know is that computation in a neural network is distributed across many nodes connected by links. To tune the behavior of the network as a whole, you need to tune each of the nodes and links, but how do you know how any one node or link contributed to the final answer? It’s complicated, and each one depends on many others. We call that “credit assignment.”

I think both brains and artificial neural networks (ANNs) need to solve the credit assignment problem. For ANNs, an algorithm called “back propagation” or just “backprop” is the industry standard solution, and it works very well. I think what brains do is different.

There are two big reasons I say this. The first issue is how work is split between nodes and links.

In ANNs, the nodes themselves are trivial, and they’re completely homogeneous across a full layer of the network, if not all the layers. Any deeper computation is about how the nodes are wired together. That is, the program is in the links (synapse weights), not the nodes.

By comparison, brain cells are both complex and diverse. We don’t know how much of the computation happens within cells vs. between them. We’re just starting to figure out what all the different kinds of cells are, but have little idea of what they’re doing. It’s clear that individual neurons do a lot, and that ensembles of cells manage each other in complex ways.

I worry saying the brain “does backprop” implies a network of trivial nodes, where tuning weight vectors is the place where learning happens. That’s likely wrong, and it obscures other possibilities.

Backprop is an algorithm I run to optimize an ANN. It needs a top-down view of the network topology and the weights of all synapses. It solves the credit-assignment problem in a clever way, usually based on the error rate compared to a known target. Then it simultaneously updates all the link weights in the network based on how the ANN responded as a whole. First you train your network, then you can use it, but not both at once.

Rather than being tuned by some external actor, brain cells manage their own relationships with their neighbors. They grow, prune, and modulate their synapses, and they decide when and how to do that based on imperfect feedback, limited information, and evolved heuristics. Brains track and minimize errors, but the targets are internally generated. This is happening continuously, with fluid transitions between acting in the real world, imagining, thinking, and learning.

I’d argue what the brain does is much harder, and much more interesting.

The Brain’s “Boss”

(this post’s image is a cross stitch I made from a pattern by Studio Ansitru. The phrase “don’t be a prick” is surrounded by a variety of cacti in pretty greens and oranges on a light blue background. In my home, it serves as a reminder to myself and to my guests.)

A popular metaphor for the mind is a pilot sitting in a cockpit, monitoring the senses and making every decision. This is obviously nonsense, but it’s an intuitive and helpful metaphor at times. The brain really does have “an executive” that thinks and plans and makes decisions, it just doesn’t have “a mind of its own,” and its power over the self is surprisingly limited. Self-control and -awareness are important for living a good life, being productive, and making ethical decisions. Unfortunately, when these faculties fail, as they often do, it’s easy to blame yourself. I find it helps to understand how these systems work, so I can set more realistic expectations for myself, which makes me less disappointed when things go awry.

Perhaps the most important thing to know about the brain is that it’s not one, unified thing. Brains are modular, with many distinct regions specialized for different tasks. Each region has different inputs and outputs, meaning they each monitor the world in different ways from different perspectives, and can cause different behaviors. Observations, insights, desires, and actions can originate in pretty much any part of the brain. All of these different regions operate together at the same time, and conscious experience is an integration of all that activity. What I think, feel, believe, and do is mostly the product of what specific regions in my brain activate together, and in what order.

This story of how brains work is surprisingly consistent across the animal kingdom. Even very simple creatures, like honey bees for instance, have complex brains with specialized regions and global integration that likely creates a sort of conscious awareness. It seems likely, however, that bees lack the sort of self-awareness and top-down control that humans do. Without it, the brain is more chaotic. Each part tries to do the right thing more or less independently. Integration means that most faculties are aware of each other and can influence each other, so there is some coordination and consistency. But focus is mostly determined by which brain region is loudest, and decisions happen moment by moment, without planning or intentional coherence.

Decentralized brains work extremely well, but they have their limitations. More complex animals have more versatile behaviors that need more explicit coordination to generate coherent, reliable, and goal-directed behavior. Most large animals seem to have this ability. In all mammals, it’s more or less identified with a brain region known as the prefrontal cortex, or PFC. The PFC is responsible for monitoring all the activity in the brain. It builds up a rich model of the self, its relationships, needs, and long-term goals. It’s responsible for planning and for exerting control over other parts of the brain. It can shut out distractions and use willpower to encourage good behaviors and discourage bad ones, even when I’d selfishly prefer to do something else.

The PFC tells the story of “you,” and has strong opinions about how that story is supposed to go.

Although all mammals have a PFC (and many other species have something analogous), the relative size of the PFC varies quite a bit between species, and that seems to correlate with executive control and what people often think of as “intelligence.” Animals with large PFCs are better at self-control, problem solving, and forming complex social relationships. Humans have exceptionally large PFCs, which partly explains why we’re so different from other species. It’s important to remember, though, this is a difference in magnitude, not in kind. It’s likely that every mammal and many other species have human-like self-awareness and self-control. It’s just a weaker faculty for them, one that acts less often, and is less able to dominate the rest of the mind.

The PFC is the closest thing to a “pilot” the brain has, but it’s better to think of it as just one brain region among many. It’s one voice in a chorus. It tends to be more bossy, spending a lot of energy trying to influence or even override other parts of the brain, but it’s not “in control.” Like a corporate executive, the PFC has only limited visibility into what the rest of the brain is doing, can’t afford to stay ever-vigilant, and can’t force a brain region to fall in line, especially when the orders run counter to that region’s nature. The PFC also isn’t capable of doing much on its own. It depends on the rest of the brain to notice things, interpret them, suggest actions, and implement them. All it can do is adjudicate and coordinate. It resolves conflict, makes plans, and advises each brain region about when and how to do its thing.

One consequence of all this is that self-control is actually a very limited and fragile thing. Often, my PFC just sits back and lets the rest of the brain work with minimal intervention. Usually that works great, but sometimes it means I miss something important. I may act out of impulse or habit and not notice until it’s too late that I’m going against my values, intentions, or best interests. Other times I know exactly what I should do, but can’t seem to make it happen. I feel unmotivated and uninspired. I can’t force myself to sit down, focus, and avoid distractions. Perhaps I’m grumpy, tired, or impatient and I do something rude or inappropriate without meaning to at all.

When this happens, it’s easy to blame myself. I lost control. I did something foolish. I acted selfishly and impulsively, like a bad person. In reality, though, this happens all the time, usually for mundane reasons that I have little control over. The PFC is energy intensive, so it gets impaired whenever my blood sugar is low, I’m tired, or I’m stressed. Other brain regions also have the ability to interfere with the PFC, especially the limbic system which manages emotions and the fight-or-flight response. Some diseases (like depression and long COVID) can cause “brain fog,” which is closely related to reduced executive function. It’s also possible to injure the PFC, from a stroke, a tumor, or a physical injury (as in the famous case of Phineas Gage).

Knowing this helps me feel a little less personally responsible when I have a lapse of self-control. It really is inevitable, common, and completely natural. Still, I want to have good judgment and do the right thing as much as possible! How do I do that? One answer is to simply be aware of my limitations and to work around them. I try to notice when I’m hungry, tired, or emotional and avoid making big decisions or socializing at those times. Instead, I might get a snack, take a break, or sleep on it so I feel more in control. The only other good technique I know is mindfulness meditation. Despite the mystical reputation, the main purpose of meditation is quite practical: it trains the PFC. By practicing the skill of observing the mind and exerting influence over it, I can build that muscle. It’s not a silver bullet, but meditation helps me use my PFC more often and more effectively, and it makes me more aware of when my PFC is in a weakened state.

So, in a sense, there really is a “pilot” in every brain. It’s just not an ever vigilant, intelligent, wise, and rational person. Instead, it’s one brain region out of many, with limited visibility and a narrow job description. The PFC observes what the other brain regions are seeing, thinking, and doing and uses that top-down view to nudge them into more coherent and effective patterns of behavior. It’s not “the self” and it’s not “in control.” In fact, it has very limited influence, isn’t always active, and even with training there are lots of common reasons it might grow weak or misbehave. For this reason, having great self-control is often less about will power in the moment, and more about avoiding temptation in the first place.

What do you think? Does this agree with your first-hand experience? Do you have any insights you could share about self-awareness, self-control, or motivation? What about being kind to yourself when your self-control inevitably falls short? If so, I’d love to hear from you in the comments.

Emotions and Cognitive Resonance

Emotions are a challenging topic. I’ve spent years learning to recognize and manage my own emotions in a productive way. I’ve also done a lot of reading, learning from psychology, cognitive neuroscience, and behavioral biology how emotions work on a theoretical and practical level. That’s useful, but the actual experience of emotions and how they shape our minds, our well-being, and our lives is incredibly subtle and complicated. I don’t think I’ll ever fully understand it, but I find it fascinating because emotions serve as a sort of interface between two different kinds of human intelligence: the body and the mind.

I like to think of the body as an autonomous robot. It does its best to survive, sustain itself, and react appropriately to whatever’s happening, recruiting the mind to understand the world and make wise decisions. That robot has evolved many different modes of behavior, each suited to particular needs, problems, and contexts. One of the main challenges for controlling the body, then, is choosing which state to be in right now. Should I look for food? Flirt with a potential mate? Take shelter and conserve my energy? Run for my life? Being successful means constantly monitoring my situation, dynamically switching from state to state, to make sure my behaviors fit the moment.

In human beings, there are tons of these “states,” but they’re hard to count since they’re so multi-dimensional and blurred at the edges. Some of the most notable ones are the core emotions (happiness, sadness, fear, and anger), but others are directly tied to the body’s function, like hunger, arousal, and fatigue. The state of the body is decided by a cocktail of hormones and neurotransmitters in the bloodstream, which coordinate activity across all the organ systems. They evolved before the brain, which is why the mind has only an indirect awareness of them. It experiences them as “feelings,” often by interpreting subtle and ambiguous signals like a flush of heat, a stirring in the guts, or a racing pulse.

That said, the brain plays a crucial role in emotions. The mind interprets what’s going on in the world to help trigger the right emotions, and emotions reconfigure the brain to serve in whatever activity the body is doing. Those chemicals in my bloodstream tune parts of my brain to be more or less active, reshaping my perception, judgment, and impulses. They shift my patterns of thought and behavior dramatically, whether I like it or not. When I’m angry, I’m more likely to perceive someone as a threat. I’m less likely to stop and think, and more likely to lash out. Looking back at the situation later with a clear head (that is, one that’s not flooded with neurotransmitters), I might see things very differently. That’s part of why recognizing emotions is so hard. Noticing and understanding are cognitive processes, happening inside my mind, while it’s being warped or even impaired by those emotions.

Another important factor at play here is how non-linear brains are. They’re collections of many different special-purpose sub-networks. These all work together in concert, combining their efforts, calling on each other, and riffing off each other to produce my stream of consciousness. This is an extremely powerful tool for creative thinking, quick intuitive action, and using metaphor to recall relevant experiences from the past. The flip side is that causality becomes very muddled. Ideas flow into one another in a sort of free-associative cascade, often forming self-reinforcing cycles. Every thought in my head is simultaneously cause and effect.

All these brain networks, ideas, and memories are linked together by associations, which cause them to resonate with each other and activate at the same time. These associations are often based on similarity or relevance, but emotions make some of the strongest and most common links. Most thoughts have emotional significance, and thinking those thoughts will evoke the associated feelings. Similarly, feelings evoke associated thoughts. More subtly, emotions also change my sense of salience. When I’m mad, I’m more likely to notice and fixate on thoughts and observations that resonate with that feeling of anger, while others get ignored.

This explains why it’s so hard to identify why I’m feeling an emotion. As a human being, I have a body evolved to react to immediate threats and opportunities, and a mind that spends much of its time thinking about abstract concepts, world events, and possible futures. This makes it very easy to misattribute my emotions. When I notice a feeling, I tend to associate it with whatever I was thinking or experiencing at that moment. This is often wrong, and that can cause problems.

A perfect example is displaced aggression. When my father was diagnosed with cancer, I was suddenly faced with many intense emotions, like anger, sadness, and fear. I carried these with me all the time, but there was little I could do to help, so I tried not to dwell on them. Nonetheless, I was much more irritable than usual. It was like my personality changed, and in a sense, it really did. I’d overreact to small slights and setbacks. I was more critical, aggressive, and impatient. I’d fixate on some insignificant detail from an email or meeting to the point where I’d be fuming about it when I got home. I managed to stay professional and respectful at work, but it took a real effort to do so.

It’s worth emphasizing how harmful misattributing emotions can be. When I’m mad, anyone who crosses my path may stir that anger just by chance, tempting me to lash out. This means the people I interact with most—my coworkers, family, and friends—are the ones most at risk. There’s also the harm done to myself. When my dad was sick, I managed not to worry about that all the time, but mostly by worrying about other things instead. The distractions were easier to face, but also emotional and not really more productive. They did nothing to relieve my stress and irritability, which lingered on until I faced the root cause and processed those emotions.

Understanding all this has led me to a practice that I find works really well. Sometimes an idea or an experience gets me suddenly riled up and emotional. Often it’s something small and petty, or something grand and abstract, both good signs that the emotion is an overreaction. Whatever thought I’m having stirs up strong feelings, which in turn drives me to obsess more on the thought in an escalating cycle. I find myself ranting or ruminating. When I notice that happening, I try to calm myself down, get some exercise, and take a break to let the neurotransmitters dissipate so I can think clearly. Then, I can use the following technique.

I set aside the idea that bothered me, and instead I focus on the feelings that idea stirs up. I sit with them for a moment, and then I look at what other ideas resonate with them. Usually, there are several. Some are big, some small. Some are immediate and concrete, some distant and abstract. Often there’s one that stands out among all the others as the most salient, and it isn’t necessarily the idea I started with. That’s when I think, “Oh! So that’s what this is really about.” In other cases, I find lots of little things, unrelated, but piling on all at once. This helps me realize there’s no one cause for how I’m feeling right now, and it’s not fair or useful to blame a scapegoat for my generally bad day.

Emotions are a critical part of how the mind works. They define human values, shape our activities, and motivate everything we do. I hope reflecting on my own experience helps illuminate that. How does it resonate with you? Is your experience similar or different? Have you noticed other quirks about your emotions, your thoughts, and how they interact? Computers are generally emotionless, but a few chat bots simulate human-like emotion, and some agent-based AIs have their own system of states and “feelings,” suited to their artificial task and environment. How do you feel about emotional AIs? I’d love to hear from you in the comments.