How Do Human Brains Function? | Andy Clark

Andy, I know you've been focusing on this concept of a predictive brain, and it certainly makes sense. I like to play table tennis, and I'm learning to hit a consistent backhand hard, and I make a lot of errors, and I'm judging it, and I don't know quite how I'm doing it, but I'm improving. It takes a long time, but I can feel that. And so it's very natural that predictive brains and error correction make sense. But what you've done is you've elevated this, not just as a function of the brain, but maybe the core idea on how the whole brain works, which is very different. So how do you justify that? Yeah. I think that's right, that the sort of provocative claim, which is at the edge of the current science and philosophy here, is that this is a fundamental operating principle of the brain, that prediction error minimization is what brains are basically about. I think a good place to start to motivate the story, but it won't be a good place to end, but we must start there, is just perception. What is it to perceive a structured external world? The claim here, which goes back to work by Richard Gregory, perhaps work by Helmholtz, work in analysis by synthesis in the psychological tradition, the thought is that when we perceive the world, the content of our percept is very much being delivered by the way that multilevel systems in the brain are trying to construct that percept for themselves using what's known about the world. This is trying to meet the sensory input with that prediction. The place where you see this is, for example, in something like the hollow mask illusion, an example that Richard Gregory loved to use. This is where you have an ordinary sort of Halloween mask or something like that. You light it from behind with the concave face looking towards you, as it were, and from the right sort of distance, that will look to you as if the face is looking outwards, as if it's convex with the nose sticking out. The kind of idea here is that this is because your brain can't help but, if you like, predict that noses are the sort of things that are going to stick out from that sort of configuration. Those predictions are actually swamping some bits of the sensory evidence, because you've got perfectly good sensory evidence coming in at your periphery saying, that's a hollow mask. In fact, it's being swamped by top-down prediction, which is bringing the ecologically most probable stimulus into view. You might think that, well, that's just a sort of a weird little exotic side case. The claim in the predictive brain literature is that this is what perception is all about. I think we see this very much if we think about something like getting to grips with sine wave speech. If anyone hasn't heard sine wave speech, lots of exemplars out there on the web, just type it in and listen to it. Sine wave speech is normal speech, stripped of a lot of the acoustic characteristics it would normally have. So it really sounds like a set of tone whistles, a bit like something a little bit like that. If you were to hear one of those speech, sine wave speech segments, and then you hear the real sentence, then you hear the sine wave speech again, you can begin to hear the real sentence through the noise. After you've done a number of sentences like that, you begin to become a sort of native speaker of sine wave speech, and you can do new sentences. I think this is what it's like to perceive in general. This is just a way of dramatizing for ourselves what must be going on when I perceive you. The fact that I managed to perceive you sitting in front of me with a certain kind of sweater on, certain kind of facial structure, is because I know about stuff like that, and these are the predictions that best accommodate the sensory information up to my expected levels of noise. That does mean that in many ways I'm kind of projecting my own expectations onto the world. So a phrase that is used here that I kind of like and I kind of don't like it, is that perception is controlled hallucination. You get this a lot in say, I think Ramesh Jain said it, Max Klaus, the machine learning person said it, and the idea is of course that your brain is forever trying to guess at what's out there, and to the extent that the guess accommodates the sensory information, then you perceive the world. One of the interesting anatomical facts that back that up is, and use the phrase top down, that there are efferent fibers from higher levels in the brain to the sensories, not just afferent fibers back to them. In some cases there are more efferent fibers from sensory organs than there are afferent ones, which without the predictive brain wouldn't make any sense, just biologically. I think that's exactly right. Even now if you open a lot of neuroscience textbooks, you will see most of the arrows going inwards in the sort of diagrams of perception, and yet about 80% of the connectivity is going downwards, and that is, you know, it would be very strange if nature had somehow spent its money on 80% downward connectivity without something like a predictive brain story being correct. And how does the top down work, because the analogy used is a visual one, so let's use the visual system. Where does it come from, and what is the feedback loops? Yeah, so there's a, top down is a kind of a hostage to fortune here, I think, because it makes people ask where the top is. You kind of think, oh, top down, so maybe I'm sitting at the very top here, and I'm the kind of little homunculus at the top doing all the predicting, which wouldn't really solve our problems. So I think the way to think about top down here is more like middle out, something like that. So imagine that there are many, many pathways leading in from the world, bringing information into the brain, and as you get further away from those impinging surfaces, more synapses away, you approach something more like the center of a sphere. That's a better model. It's not a step ladder, it's the center of a sphere. And what level in all of that processing is currently at the top, I think is something which, in the sophisticated versions of these stories, varies according to the task that you're trying to perform. So one of the things that predictive brains have to do is moment by moment estimate where the good information is relative to what they're trying to do, or what the organism is trying to do, or what the extended organism is trying to do. The predictive brain certainly makes sense in perception, and your claim is that this is the core idea. Certainly it is critical whether its core idea is an aggressive claim, which a lot of people are taking seriously. Are there other aspects of cognition other than perception which the predictive brain also makes sense in internal rational thinking, or emotion, or other aspects besides pure perception? Yeah. I think a lot of the interest of these stories comes from the fact that they are madly ambitious. The current most ambitious versions are probably coming out of the work of Carl Friston and colleagues at University College London, and they claim that this is a unified story about perception, cognition, action, emotion. Anything you can think of that falls broadly under that umbrella of what minds do, it turns out that there's an elegant story to be told in terms of the predictive brain where a lot of stuff falls into place. So if you think about cognition, first of all, I think it's fairly easy to see that as soon as you're in the business of trying to predict the flow of sensory inputs, you're kind of in the business of knowing about your world. If I know how this scene will evolve over many scales of space and time, I know that that little rustle of the curtain presages my cat coming in, my cat is likely to meow and ask for food or something, I'm kind of understanding my world. So I think this is a vision of perception which makes it continuous with cognition. It also makes it continuous with imagination because if you can do this trick of constructing the sensory flow from the top down, then you can construct a sort of approximation to the sensory input when there's no sensory input. It's the same kind of process. So now we've got perception, cognition and imagination all coming on board at the same time. There's also an extension to action and an extension to emotion. So what you're, in essence, saying is that these divisions of mentality that we have created, the difference between perception and cognition, for example, or cognition and imagination, may blur because sitting below them all is a fundamental principle of a predictive brain and error correction in terms of a directionality. And so if you have that principle, it's sort of almost as a unifying principle, like in particle physics you have higher and higher energy to unite the forces. So here your high energy is the predictive brain and so that unifies these heretofore separate categories of perception, cognition, etc. Does that make sense? That's exactly the attractiveness and also the scariness of the story. So I think the big unifying stories here sit at sort of the level of something like string theory. So, you know, if you were really looking for a way to unify everything about particle physics, maybe you need something a bit exotic like that. Currently these stories are about that sort of level, I think. It's not quite obvious what it takes to disprove them, not quite obvious what it takes to sort of show that we should be adopting the unified story rather than just some bits of the unified story and lots of other principles too. But at the same time, you know, if I was to explore this in the case of, say, emotion, there would be some very interesting things to say. The same for action. So action here comes out as proprioceptive prediction. So if I'm going to move my hand in a certain way, I predict the way that proprioceptive information about the orientation of my limb in space and sort of muscle and tendon properties, I predict the way that that information would flow. And this time I get rid of the prediction error by actually bringing that flow about. So it's a self-fulfilling prophecy. This is very elegant, very neat way of doing motor control. And there are stories to tell about emotion too. So at this point, I think we've got a pretty big but interesting account of what minds and brains are.