Instant Genius - How AI is helping us to understand how our brains learn

Episode Date: March 3, 2025

The human brain is surely one of the most complex and fascinating creations in all of nature. It’s through them that we experience the world around us, develop knowledge and skills, and carry within us memories of events and experiences that occurred decades in the past, which we can later draw on to face new challenges. But when it comes to understanding exactly what is going on between our ears, we’re only just starting to scratch the surface. Could artificial intelligence technology help us in this quest? In this episode, we’re joined by Prof Andrew Saxe, a computational neuroscientist based at University College London. He tells us about the work his lab is carrying out on modelling the function of our brains using artificial intelligence, what this research has already uncovered about how our brains learn and what we can look forward to discovering in the coming decade or so. Andrew is a finalist for the Blavatnik Awards for Young Scientists.   https://blavatnikawards.org/news/items/finalists-announced-2025-uk-blavatnik-awards-young-scientists/ Learn more about your ad choices. Visit podcastchoices.com/adchoices

Ambition comes in all shapes and sizes. At First Citizens Bank, we roll with your goals, because we're built for what you're building. Fit for your ambition, First Citizens Bank. Peak pollination season, and my business is scaling fast. To keep the nectar flowing, I need a phone plan with top priority data speeds. That's why I chose GoogleFi wireless. My connections stay strong even when the hive is buzzing.

Plus, unlimited plans started $35 a month. Now that's a new. deal that doesn't stay. Explore GoogleFi Wireless plans today. Plus taxes and government fees. GoogleFi Wireless is not subject to data traffic deprioritization during times of high network usage. You said this place was steps from the water.

We just haven't found the steps yet. How much did we save? Enough. Enough to get lost. Or you could book a stay with Hilton. Welcome to your ocean front room. Just steps from the water. The Hilton sale is on.

now. Book on Hilton.com or the Hilton app and save up to 20% to get the stay you expected. When you want savings, not surprises. It matters where you stay. Hilton, for this day. This podcast is sponsored by name, audio and focal. Streaming has made music more accessible than ever, but true listening is about more than ease. It's about quality. British audio experts name Audio, alongside French acoustic specialist focal, combine handcrafted tradition with cutting-edge innovation and high-end materials, delivering digital precision with analogue warmth. So you can experience exceptional sound at home. Music just as the artist intended. Visit name audio.com to learn more. Hello and welcome to Instant Genius, a bite-sized master class in podcast form. Every Monday and Friday

you will hear world-leading scientists and experts talking about the most fascinating ideas in science and technology today. I'm Jason Goodyear, commissioning editor, a BBC science focus. The human brain is surely one of the most complex and fascinating creations in all of nature. It's through them that we experience the world around us, develop knowledge and skills, and carry within us memories of events and experiences that occurred decades in the past, which we can later draw on to face new challenges. But when it comes to understanding exactly what's going on between areas. We're only just starting to scratch the surface. So could artificial intelligence technology help us in this quest? In this episode, we're joined by Professor Andrew Sacks,

a computational neuroscientist based at University College London. He tells us about the work his lab is carrying out and modelling the function of our brains using artificial intelligence, what this research has already uncovered about how our brains learn, and what we can look forward to discovering in the coming decade or so. So welcome to the podcast. Thanks very much for joining us. It's a pleasure. Thanks for having me. So first off, I think a good place to start is, can you tell us a bit about your background? You know, you work in a very interesting area. Yeah. And most people in this area came from some, I think, unusual starting place because it wasn't really a, you know, it's not a fully formed field. So my path took me through electrical engineering. But I studied a lot of neuroscience and machine learning.

and computer science as well. And now I also spend a lot of time with physicists. And the goal is to understand how the brain works somehow. So, yeah. So your specific area is concentrated on learning. That's right. So can you give us a sort of Cliff's Notes version of that? And then we'll get into the sort of meat.

Right. Yeah. I find learning fascinating because if you think of how many things you do that you learned, I mean, it's most, it was a huge chunk of our mental. lives is something we've had to learn. I mean, we live in worlds that evolution hasn't seen, right? We have skyscrapers staring at phones all the time. So these are all things we've had to learn. And I think for understanding the mind and ourselves, if you have a theory of learning, you might manage to explain a lot of different things with that one theory. So in the same way that

in physics, Newton's laws can explain many different physical situations. A theory of learning might make how our brain processes sound or how it processes vision all look sort of the same, how do you learn from these different signals? And so your work is centered on artificial intelligence. So exactly how do we use artificial intelligence to learn about learning? Yeah, it's not clear if the way that our AI systems work is anything like how the brain works. But they were inspired by it. So Jeff Hinton, one of the key figures in developing artificial intelligence systems you might

have used, like chatbots, for instance, is very interested and deeply concerned with how the brain works. And there's been a shared history between these fields. And so the AI systems we use are called artificial neural networks because they resemble some aspects of brain networks, not all, but some. And now that they're working quite well, the question is, can we we learn anything about how our brains work? Is it more than inspiration? And a lot of what my lab is doing is trying to test that question. Can we do experiments that would indicate the brain works something like these artificial systems or reject that hypothesis? And then we can move on to try to find more brain-like learning methods. So what exactly is a neural network? And how do you go about

creating one? In the brain, you have cells which signal each other with little electrical pulses. And they form a large network. And the intriguing thing about them is each one is reasonably simple, right? It's just a little cell. But once you hook them all into a network, apparently, as far as we can tell, that allows you to do everything you do. And so an artificial neural network is trying to get at that idea that you have many simple computational elements interconnected into a big network.

And then that network is able to do something. something quite sophisticated, like chat to you in natural language. So the individual elements, we model them mathematically, the very simple operation. And then the whole network, the connections between them get tuned by learning. So that's where the learning comes in. It's in which connections get strengthened or weakened. So is this something that's called deep learning? Yeah, exactly. Yeah, so what is that? Can you explain that for us? you could imagine a system that receives an input and maps directly to an output.

So maybe there's light falling on your eye. And it could be that the nerve cells go straight to a muscle and cause a muscle to move in some way. And it turns out if you, at that level, it's very challenging to map from input to output in one step because that mapping is going to have to be so complex. I think of all the different things you have to do in response to light patterns. So instead, basically, we divide and conquer. We split it up into many, many layers. That's the deep aspect.

And each layer doesn't have to do that much. It just gets a little closer to the right transformation. But once you stack them all together, then it's doing something quite impressive. There's a core idea behind this whole area is somehow very simple elements. When you configure them into a complicated network architecture, can do way more than what you might have thought from how simple each one is. And depth is one aspect of that. So let's have a look at some specifics of your research then.

So I've been having a look, and it seems that you broadly split it into three prongs. So the first of which you call learning in complex architectures. So what do you mean by that? Yeah, a classical deep network in the machine learning world is just a feed-forward chain. One layer goes to the next, goes to the next, goes to the next. But if you look inside the brain, you don't see that. You see much more intricate connectivity. So there is some of that.

In the visual system, for instance, you have early primary visual cortex, which selected for edges and images, and then secondary visual cortex and so forth. So there's some amount of stage-like processing. But there's also looping connections and connections that skip levels. And that's clearly important. Recurrent connections that neurons that connect back onto the. themselves. And so we have to get a handle on what those are contributing. So that's what I mean by more complex architectures. So another one that you split into is learning with prior knowledge.

So this is really interesting. How on earth do you do that with a neural network? Yeah, we operationalize it as just what the network already knew before you continued training it. So maybe you've trained it to be a chatbot. But then after that, you're trying to train it to also solve math problems. And some of what it's learned might help it in that task, and some of what it's learned might hinder it. And that's the interesting question. Of course, as humans, almost everything we learn, we're learning amidst the background of all of the prior knowledge we've learned across our whole lifespan. So that's a key factor you have to include in these theories. So you mentioned they're training these systems. So how do you go about that?

It's simple when it comes down to it. You basically have a data set of inputs and target outputs that you want it to produce. So maybe it's an image of cats and dogs. And you have labels, whether that image contains a cat or dog. You have a big database of these, many millions. And you show an image to your artificial network and you let the activity flow through. And maybe you show it a cat, but initially it says, I think that's a dog. So it's wrong.

And then what you do is you adjust all of the connections in that network, just a little teeny bit, very small bit, to make it slightly more correct the next time. So maybe if it saw that image again, it would say, that's a cat, or at least be less certain it was a dog. And you do that small tweak after one presentation of an example. You do that millions of times. And then eventually the network settles on a configuration where it's mostly getting things correct. So how long does that take? it can't take quite a while. And, you know, in terms of time on the wall, a wall clock time,

you can use computers and various parallel processing systems to make it go faster, but you're still taking a ton of energy. So, yeah, the speed of training these systems and the size of the datasets they have to train on for the chatbots, it's a sizable fraction of the internet. That is a major challenge and something that these theories hopefully will eventually address is how can we speed up learning? Because, of course, for humans, we do take a while to learn, but we don't take nearly as much energy as these systems. So what sort of hardware do you use? Is it like really specialist kit? Yeah, although now I suppose maybe it's becoming very common. So there is a split in the field historically between a serial computer, which can only do one thing at a time,

and it's processing, that's what we classically think of as a computer. One piece of information at time is considered, and we got those to be quite fast, so serial computer can, you know, rapidly go through a bunch of steps. But if you have a giant network, it's much more convenient to be able to compute lots of things in parallel, because these simple computing elements, they don't need to know what any of the others are doing, so they can just do their own operation. And rather than considering each one in sequence with what you have to do with a serial computer, you could just do them all in parallel in graphics processing units. GPUs is what mainly are used, which were repurposed from computer vision to machine learning.

Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the Unreal College deal, everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 premium and a year of Xbox GamePass Ultimate with a custom color Xbox wireless controller. at Windows.com slash student offer. While supplies last, ends June 30th, turns at AKA.m.m.S. slash college PC.

When you need to build up your team to handle the growing chaos at work, use Indeed-sponsored jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75-sponsored job credit

at Indeed.com slash podcast. That's Indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed sponsored jobs. This podcast is sponsored by Name, Audio and Focal. With over 100 years of combined expertise, Name and Focal have been bringing music to listeners

just as the artist intended. Since day one, this mantra has shaped every innovation in high-fi design, technology, and acoustic engineering, balancing craftsmanship and tradition with pioneering thinking. Name Audio pushes cutting-edge technology to ensure digital precision whilst sustaining Pratt, pace, rhythm and timing, the elusive quality that makes music feel alive and gives it emotional texture. Today, in partnership with French acoustic specialist's focal, name audio creates systems that deliver exceptional sound,

and unforgettable listening experiences at home. for yourself at a focal powered by name boutique. Visit focal powered by name.com for more information. So another sort of topic that you bring up is something you call learning with non-linearity and context. So that says a little bit complicated. Can we break down exactly what that means? It is complicated. If you think of a real world task, there are some tasks that are linear, by which I mean that the input output app is quite simple. Maybe you're asking, how housing price varies with house size. And okay, that's going to be roughly increasing. No surprises there. But most tasks in our world are nonlinear. It's going to be far more complicated. And even

housing price might depend on which markets you're in. So certain contextual factors or whether you think there's likely to be floods and so forth. And so when you get to real world problems, you have to deal with the non-linearity of the world. And this is a big challenge for theory. So For theory, we have good methods for analyzing linear systems, much harder to analyze nonlinear systems. But that's really what we want an explanation of. So, sort of sticking with humans then, so one of the things that helps us learn is our memories. So this is something you also study, isn't it, how sort of long-term memories are formed?

Yes, and there's an interesting feedback loop between the machine learning systems and the models of human memory here. So the way the AI systems are trained is you have a database of images, which you can think of as a memory, right? The database has to be stored, and then you play it to this artificial network. And in the brain, there's something of a similar situation where you have a brain region called the hippocampus, which is thought to be an episodic memory store. So it just memorizes episodes that happen to you. And then at various times, like when you're resting or when you're sleeping, those episodes can be replayed and broadcast to the rest of your cortical network. And, of course, a long-standing idea has been that that's to train the cortical network.

And so we offered a theory built on these deep learning models of how this interaction might facilitate generalization. So the whole system is orchestrated to extract useful patterns that are out there in the world that allow animals and humans to generalize to new experiences. So can you sort of summarize some of your headline results that you've had so far? Sure. So, well, in the context of memory, there's been a long-running debate about whether hippocampus stores memories permanently or whether eventually it can train them into cortex. And this is, you know, there's many, many, many experimental papers about this. And it's an important topic because in Alzheimer's, hippocampus is the brain area that

is targeted and lots of memory diseases relate to these interactions. We do want to understand it. So this decades-old puzzle, we think we have at least a new perspective on. And it is that if you try to replay information from hippocampus to cortex, sometimes that information is too idiosyncratic to find generalizable structure. In the machine learning system, sometimes when you keep training it, it actually gets worse. And because of that, you should not continue training the system at that point. And if you can't continue training the system, then those memories would have to stay in the hippocampus forever if you wanted to memorize them. And so this is a new principle for when memories reside in different parts of the brain. It basically says, to what extent can you generalize them?

And if memories are very idiosyncratic, they might have to forever stay in your hippocampus. And we think this explains some of this discrepancies in the literature. So obviously this is all sort of super interesting from a kind of pure science viewpoint. But you mentioned there Alzheimer's. So what sort of potential applications can come out of this research? Yeah, it's a bit hard to say. And I should start by saying that what my work has done is proposed theories, which are pretty much untested. So we think they explain a lot of existing data, but we really need new experiments to pin them down.

So they are, this is fresh theorizing, right? That's one comment. But supposing they were correct, proven correct in the future, then it could be that pathologies with this mechanism of trying to regulate generalization, for instance, might be the cause of several diseases. So for instance, post-traumatic stress disorder could be maladaptively holding on. to a memory, which actually doesn't generalize, but you think it does. So you were in a very traumatic situation, and you think that now that danger applies broadly, it's generalized, whereas in fact, it should have been restricted. So it's too early to give a specific application, but that's the

hope. And maybe a broader comment on that is when you understand something is usually when you can control it. So that's why we do basic science. For me, it's not only just the interest of It is that once you get that explanatory theory, that's when you'll be able to control the system for whatever end you're interested in. So can we use these same sort of systems to do sort of dry runs of potential treatments? I mean, is that a thing? I hope it will be a thing. I haven't done that, but that's certainly the goal.

That is absolutely the goal. And there are some seeds of that approach. So Chris Pack at McGill, for instance, has had some success developing treatments for stroke that are very much about how to make people learn effectively in the recovery period. So a theory of learning can help for rehabilitation, of course. Also, so there's additional psychiatric, basically the field of computational psychiatry is trying to use these models to understand how you can improve psychiatric treatments. So sort of looking forwards, I mean, everybody knows that artificial intelligence is moving at just an incredible pace.

Is it the same for this sort of research? Is it sort of every couple of months? We've got something new coming out. Well, every couple of months, there's certainly new things coming out. But I'm not sure the pace is quite the same because when you're building a system, the artificial intelligence challenge is largely engineering challenge. Once it works, you get to celebrate. That's it.

It works. What we're trying to do is also understand. And theory often lags practice. I think historically it often has lagged practice. You think of we could build cathedrals and temples around the world. They're very impressive buildings without an extensive knowledge of physics. But once we did have our physical knowledge, that's when you can build skyscrapers or bridges that span large bodies of water or send an astronaut to the moon. So you want that knowledge, but it often does come afterwards.

Yeah, a lot of people talk about the singularity. But is there a tipping point with this sort of thing where we're like, whoa, you know, now this is the target that we're all aiming for. And once we reach that, you know, things will really start getting going. Yeah, I mean, on the AI side, it could be. So one possibility is we'll have AI systems that are better theorists than I am. And then I can go do something else, right? Learn to surf. I think that's a possibility. It is a possibility. We'll just have to see. I think right now I still think there's a lot of important theoretical work to do. So I'm not sure we're going to get there. And we also don't know the relationship between these systems and the brain. And I am interested, I think, as We all are really and what we are as a human being. So the human aspects is important. But yeah, I think ultimately we'll just have to see where we are in 50 years. Maybe the AI systems will be better neuroscientists than us as well. So let's sort of by way of summing up then.

Let's have a look. What are you sort of working on in the next five or 10 years? What can we be looking forward to for research coming out of your lab? Yeah. So one direction is understanding reasoning capabilities. So the most recent crop of AI models have quite impressive mathematical reasoning abilities. I think that's an opportunity to understand and formalize that of some of those abilities in humans, which have been quite resistant to finding ways of thinking about.

But the current AI systems are doing that. So it would be important to understand how they're doing it. Can we improve them? Can we make them train more quickly? and what does it tell us about how brains do it. Another direction that I want to go in is empirical tests of the core learning principle. So almost all of the AI systems are trained with the same learning mechanism. Pretty much what I said, tweak the weights a little bit to reduce the error on any given example.

And we don't know if the brain can do that. It's a pretty sophisticated mathematical operation. And so I think we should turn that into an experiment and do that experiment. and then that would be a linchpin that connects artificial intelligence, psychology, and neuroscience together if that one learning principle turns out to be a good description of what's happening in the brain in certain cases. So having said that, you know, are you optimistic that we will find solutions to these problems? Oh, absolutely. But the timescale is the issue. I am very optimistic if you give me infinite time, whether it will happen in the next, you know, short.

while is another matter, but it's okay. It still will be worth it if it happens a long time from now. Thank you for listening to this episode of Instant Genius, brought to you from the team behind BBC Science Focus. That was Professor Andrew Sacks. If you liked what you just heard, then please do consider subscribing to Instant Genius on your preferred podcast platform. You can now also find us on YouTube at at Science Focus. The current issue of BBC Science Focus magazine is out of now pick up a copy wherever you buy your favourite magazines or download us on your app store of choice you can also find us on apple news or online at sciencefocus.com this podcast is sponsored by name audio and focal the texture and emotional depth of music

can be lost through digital sources or poor signal name audio believes you can have digital precision with analog warmth alongside french acoustic specialist vocal name creates a high-end audio systems combining innovation with craftsmanship so you can listen to music, just as the artist intended. Discover more at name audio.com.