The real reason for brains - Daniel Wolpert

[Music] [Applause] I’m a neuroscientist, and in neuroscience, we have to deal with many difficult questions about the brain. But I want to start with an easiest question, and the question you really should have all asked yourself at some point in your life because it’s a fundamental question if we want to understand brain function. And that is, why do we and other animals have brains? Not all species on our planet have brains, so if we want to know what the brain is for, let's think about why we evolved one.

Now, you may reason that we have one to perceive the world or to think, and that's completely wrong. If you think about this question for any length of time, it’s blindingly obvious why we have a brain. We have a brain for one reason and one reason only, and that's to produce adaptable and complex movements. There is no other reason to have a brain. Think about it: movement is the only way you have of affecting the world around you. That's not quite true; there's one other way, and that's through sweating. But apart from that, everything else goes through contractions of muscles.

So think about communication. Speech, gestures, writing, sign language—they're all mediated through contractions of your muscles. So it's really important to remember that sensory memory and cognitive processes are all important, but they're only important to either drive or suppress future movements. There can be no evolutionary advantage to laying down memories of childhood or perceiving the color of a rose if it doesn’t affect the way you're going to move later in life.

Now, for those who don't believe this argument, we have trees and grass on our planet without brains. But the clinching evidence is this animal here: the humble sea squirt. This rudimentary animal has a nervous system, swims around in the ocean in its juvenile life, and at some point in its life, it implants on a rock. The first thing it does in implanting on that rock, which it never leaves, is to digest its own brain and nervous system for food. So once you don't need to move, you don't need the luxury of that brain. This animal is often taken as an analogy to what happens in universities when professors get tenure, but that's a different issue.

So, I am a movement chauvinist. I believe movement is the most important function in the brain. Can anyone tell you that it’s not true? Now, if movement is so important, how well are we doing understanding how the brain controls movement? The answer is we’re doing extremely poorly. It’s a very hard problem. But we can look at how well we're doing by thinking about how well we're doing building machines that can do what humans can do. Think about the game of chess: how well are we doing determining what piece to move where? If you put Garry Kasparov here, when he’s not in jail, against IBM's Deep Blue, well the answer is IBM's Deep Blue will occasionally win.

I think if IBM's Deep Blue played probably anyone in this room, it would win every time. That problem is solved. What about the problem of picking up a chess piece, dexterously manipulating it, and putting it back down on the board? If you put a 5-year-old child's dexterity against the best robots of the day, the answer is very simple: the child wins easily. There’s no competition at all.

Now, why is that top problem so easy and the bottom problem so hard? One reason is a very smart five-year-old could tell you the algorithm for that top problem: look at all possible moves to the end of the game and choose the one that makes you win. So it’s a very simple algorithm. Now, of course, there are a lot of moves, but with fast computers and some approximations, we come close to the optimal solution.

When it comes to being dexterous, it’s not even clear what the algorithm is you have to solve to be dexterous. You have to both perceive and act on the world, which has a lot of problems. But let me show you cutting-edge robotics. Now a lot of robotics is very impressive, but manipulation robotics is really still in the Dark Ages. This is the end of a PhD project from one of the best robotic institutes, and the student just trained this robot to pour water into a glass. It’s a hard problem because the water sloshes about, but it can do it. However, it doesn’t do it with anything like the agility of a human.

Now if you want this robot to do a different task, that’s another 3-year PhD program. There is no generalization at all from one task to another in robotics. Now we can compare this to cutting-edge human performance. So what I'm going to show is Emily Fox winning the world record for cup stacking. Now the Americans in the audience know all about cup stacking. It’s a high school sport where you have 12 cups you have to stack and unstack against the clock in a prescribed order.

This is her getting the world record in real time, and she’s pretty happy. We have no idea what is going on inside her brain when she does that, and that's what we'd like to know. So in my group, what we try to do is reverse engineer how humans control movement. It sounds like an easy problem: you send a command down, it causes muscles to contract, your arm or body moves, and you get sensory feedback from vision, from the skin, from muscles, and so on.

The trouble is these signals are not the beautiful signals you want them to be. One thing that makes controlling movement difficult is, for example, sensory feedback is extremely noisy. Now, by noise, I do not mean sound. We use it in the engineering or neuroscience sense, meaning random noise corrupting a signal. So, in the old days, before digital radio, when you were tuning in your radio and you heard on the station you wanted to hear, that was the noise. More generally, this noise is something which corrupts the signal.

For example, if you put your hand under a table and try to localize it with your other hand, you could be off by several centimeters due to the noise in sensory feedback. Similarly, when you put motor output or movement output, it’s extremely noisy. Forget about trying to hit the bull's eye in darts; just aim for the same spot over and over again, and you have a huge spread due to movement variability.

More than that, the outside world or task is both ambiguous and variable. The teapot could be full; it could be empty; it changes over time. So we work in a whole sensory movement task soup of noise. Now, this noise is so great that society places a huge premium on those of us who can reduce the consequences of noise. So if you’re lucky enough to be able to knock a small white ball into a hole several hundred yards away using a long metal stick, our society will be willing to reward you with hundreds of millions of dollars.

Now, what I want to convince you is the brain also goes through a lot of effort to reduce the negative consequences of this sort of noise and variability. To do that, I'm going to tell you about a framework which is very popular in statistics and machine learning over the last 50 years called Bayesian decision theory. It’s more recently a unifying way to think about how the brain deals with uncertainty. The fundamental idea is you want to make inferences and then take actions.

So let's think about the inference. You want to generate beliefs about the world. So what are beliefs? Beliefs could be: Where are my arms in space? Am I looking at a cat or a fox? But we're going to represent beliefs with probabilities. So we're going to represent a belief with a number between zero and one—zero meaning I do not believe it at all, and one means I'm absolutely certain. Numbers in between give you the gray levels of uncertainty.

The key idea in Bayesian inference is you have two sources of information from which to make your inference. You have data, and data in neuroscience is sensory input. So I have sensory input which I can take in to make beliefs. But there’s another source of information, and that’s effectively prior knowledge. You accumulate knowledge throughout your life in memories. The point about Bayesian decision theory is it gives you the mathematics of the optimal way to combine your prior knowledge with sensory evidence to generate new beliefs.

I put the formula up there—I’m not going to explain to you what that formula is, but it's very beautiful, and it has real beauty and real explanatory power. What it really says is what you want to estimate is the probability of different beliefs given your sensory input. Let me give you an intuitive example. Imagine you’re playing tennis, learning to play tennis, and you want to decide where the ball is going to bounce as it comes over the net towards you. There are two sources of information. Bayes' rule tells you there’s sensory evidence you can use: visual information, auditory information, and that might tell you it’s going to land at that red spot. But you know that your senses are not perfect, and therefore there’s some variability in where it’s going to land, shown by that cloud of red representing numbers between 0.5 and maybe 0.1.

That information is available on the current shot, but there’s another source of information not available on the current shot but only available by repeatedly experiencing the game of tennis. That’s that the ball doesn’t bounce with equal probability over the court during a match. If you’re playing against a very good opponent, they may distribute it in that green area, which is the prior distribution, making it hard for you to return.

Now, both these sources of information carry important information. What Bayes' rule says is I should multiply the numbers in the red by the numbers in the green to get the numbers in the yellow, which have the ellipses, and that’s my belief. So it’s the optimal way of combining information.

Now, I wouldn’t tell you all this if it wasn’t that a few years ago we showed this is exactly what people do when they learn new movement skills. What it means is we really are Bayesian inference machines. As we go around, we learn about the statistics of the world and lay that down, but we also learn about how noisy our own sensory apparatus is and then combine those in a real Bayesian way.

A key part of the Bayesian inference is this part of the formula, and what this part really says is I have to predict the probability of different sensory feedback given my beliefs. That really means I have to make predictions of the future. I want to convince you that the brain does make predictions of the sensory feedback it’s going to get, and moreover, it profoundly changes your perceptions by what you do.

To do that, I’ll tell you about how the brain deals with sensory input. So you send a command out, and you get sensory feedback back, and that transformation is governed by the physics of your body and your sensory apparatus. But you can imagine looking inside the brain, and here’s inside the brain. You might have a little predictor, a neural simulator of the physics of your body and your senses. So as you send the movement command down, you tap a copy of that off and run it into your neural simulator to anticipate the sensory consequence of the reaction.

So, as I shake this ketchup bottle, I get some true sensory feedback as a function of time on the bottom row, and if I’ve got a good predictor, it predicts the same thing. Well, why would I bother doing that? I’m going to get the feedback anyway. Well, there are good reasons. Imagine as I shake the ketchup bottle, someone very kindly comes up to me and taps it on the back for me. Now, I get an extra source of sensory information due to the external act. So I get two sources: I get you tapping on it and I get me shaking it, but from my senses' point of view, that just combines together into one source of information.

Now, there are good reasons to believe that you would want to be able to distinguish external events from internal events because external events are actually much more behavior-relevant than feeling everything that’s going on inside my body. So one way to reconstruct that is to compare the prediction—which is only based on your movement commands—with the reality, and any discrepancy should hopefully be external. So as I go around the world, I’m making predictions of what I should get, subtracting them off; everything left over is external to me.

What evidence is there for this? Well, there’s one very clear example where a sensation generated by myself feels very different than one generated by another person. We decided the most obvious place to start was with tickling. It's been known for a long time you can’t tickle yourself as well as other people can, but it hasn’t really been shown it’s because you have a neural simulator simulating your own body and subtracting off that sense.

So we can bring the experiments in the 21st century by applying robotic technology to this problem. Effectively, what we have is someone holding a stick in one hand attached to a robot, and they're going to move that back and forth. We’re going to track that with a computer and use it to control another robot, which is going to tickle their palm with another stick. Then we’re going to ask them to rate a bunch of things, including ticklishness.

I’ll show you just one part of our study. Here, I’ve taken away the robots, but basically, people move with their right arm sinusoidally back and forth, and we replace that with the other hand with a time delay—either no time delay in which case it’s like you're just tickling your palm, or with a time delay of a tenth, two-tenths, or three-tenths of a second. So the important point here is the right hand always does the same thing: sinusoidal movement. The left hand is the same input: tickle. All we’re playing with is a temporal causality.

As we go from no time delay to 0.1 seconds, it becomes more ticklish. As you go from 0.1 to 0.2 seconds, it becomes more ticklish again, and by 0.2 seconds, it’s equivalently ticklish to if the robot just tickles you without you doing anything. So whatever is responsible for this cancellation is extremely tightly coupled to temporal causality. Based on this and other studies, we’re really convinced in the field that the brain is making precise predictions and subtracting them off from the sensations.

Now, I have to admit these are the worst studies my lab has ever run because the tickle sensation in the palm comes and goes. You need large numbers of subjects in these studies to make them significant, so we were looking for a much more objective way to assess this phenomenon. In the intervening years, I had two daughters, and one thing you know about children in the back seats of cars on long journeys: they get into fights, which starts with one of them doing something to the other.

The other one retaliates, and it quickly escalates. Children tend to get into fights which escalate in terms of force. Now, when I scream at my children to stop, sometimes they would both say to me the other person hit them harder. Now, I happen to know my children don’t lie, so I felt as a neuroscientist it was important—how could I explain how they were telling inconsistent truths?

We hypothesized, based on the tickling study, that when one child hits another, they generate the movement command, they predict the sensory consequences, and subtract it off. So they actually think they’ve hit the person less hard than they have, rather like the tickling, whereas the passive recipient doesn’t make the prediction and feels a full blow. So if they retaliate with the same force, the first person will think it’s been escalated.

We decided to test this in the lab. We don’t work with children; we don’t work with hitting, but the concept is identical. We bring in two adults and tell them they're going to play a game. Here’s player one and player two sitting opposite each other, and the game is very simple. We start off with a motor with a little lever with a little force transducer, and we use this motor to apply a force down to player one’s fingers for 3 seconds, and then it stops.

This player has been told to remember the experience of that force and use their other finger to apply the same force down to the other subject’s finger through a force transducer. They do that, and player two has been told to remember the experience of that force and use their other hand to apply the force back down. They take turns applying the force they’ve just experienced back and forth, but critically, they’re briefed about the rules of the game in separate rooms, so they don’t know the rule the other person is playing by.

What we measure is the force as a function of turns. If we start with a quarter of a Newton, there are a number of turns; the perfect would be that red line. What we see in all pairs of subjects is a 70% escalation in force on each go. So it really suggests when you're doing this—that based on the study and others we’ve done—the brain is canceling the sensory consequences and underestimating the force it's producing.

So it really shows that the brain makes predictions and fundamentally changes the percepts. So we’ve made inferences; we’ve done predictions. Now we have to generate actions. What Bayes’ rule says is given my beliefs, the action should, in some sense, be optimal. But we've got a problem: tasks are symbolic. I want to drink; I want to dance. But the movement system has to contract 600 muscles in a particular sequence. There’s a big gap between the task and the movement system, so it can be achieved in infinitely many different ways.

Think about just a point-to-point movement. I could choose these two paths out of an infinite number of paths. Having chosen a particular path, I can hold my hand on that path. There are infinitely many different joint configurations, and I can hold my arm in a particular joint configuration, either very stiff or very relaxed. So I have a huge amount of choice to make.

Now it turns out we are extremely stereotypical. We all move the same way pretty much. It turns out we’re so stereotypical that our brains have dedicated neural circuitry to decode the stereotypic. So if I take some dots and set them in motion with biological motion, your brains have circuted you to understand instantly what’s going on: the happy, the sad, the old, the young—there’s a huge amount of information.

If these dots were cars going around a racing circuit, you would have absolutely no idea what’s going on. So why is it we move the particular ways we do? Well, let's think about what really happens. Maybe we don’t all quite move the same way; maybe there’s variation in the population, and maybe those who move better than others have a better chance of getting your children into the next generation.

So, on evolutionary scales, movements get better, and perhaps throughout life, movements get better through learning. So what is it about a movement which is good or bad? Imagine I want to intercept this ball. Here are two possible paths to that ball. If I choose the left-hand path, I can work out the forces required in one of my muscles as a function of time, but there’s noise added to this. So what I actually get, based on this lovely smooth desired force, is a very noisy version.

If I play the same command through many times, I will get a different noisy version each time because noise changes each time. So what I can show you here is how the variability of the movement will evolve if I choose that way. If I choose a different way of moving on the right, for example, then I’ll have a different command, different noise playing through a nonlinear system—very complicated. All we can be sure of is the variability will be different.

If I move in this particular way, I end up with smaller variability across many movements. So if I had to choose between those two, I would choose the right one because it’s less variable. The fundamental idea is you want to plan your movement so as to minimize the negative consequence of the noise. One intuition to get is that actually the amount of noise or variability, as I show here, gets bigger as the force gets bigger. So you want to avoid big forces as one principle.

We’ve shown that using this, we can explain a huge amount of data—that exactly people are going about their lives planning movements so as to minimize negative consequences of noise. So, I hope I’ve given you the idea that the brain is there and evolved to control movement, and it’s an intellectual challenge to understand how we do that. But it is also relevant for disease and rehabilitation. There are many diseases that affect movement, and hopefully, if we understand how we control movement, we can apply that to robotic technology.

Finally, I want to remind you that when you see animals do what look like very simple tasks, the actual complexity of what’s going on inside their brain is really quite dramatic. Thank you very much.

Quick, um, quick question for you there. So, so you're, um, a movement chauvinist. Does that mean that you think that the other things that we think our brains are about—the kind of the dreaming, the yearning, the falling in love, and all these things—are a kind of, um, a sideshow, an accident?

They're not an accident. I think they’re all important to drive the right movement behavior to get reproduction in the end. So I think people who study sensation or memory without realizing what you're laying down memories of childhood—that if you get most of our childhood, for example, it’s probably fine because it doesn’t affect our movements later in life. You only need to store things which are really going to affect movement.

So you think that people thinking about the brain and consciousness generally could get real insight by saying where does movement play in this game? People have found out, for example, that studying vision in the absence of realizing why you have vision is a mistake. You have to study vision with the realization of how the movement system is going to use vision, and it uses it very differently once you think of it that way.

Well, that was quite fascinating. Thank you very much. [Applause] Indeed.