As you can tell, I am a self-confessed movement chauvinist. I believe movement is the most important function of the brain, and to understand movement requires us to understand how perception, memory and cognition affect action. However, the effortless ease with which humans move – our arms, our eyes, even our lips when we speak – masks the true complexity of the control processes involved. This is evident when we try to build machines to perform human control tasks. While computers can now beat grandmasters at chess, no computer can yet control a robot to manipulate a chess piece with the dexterity of a six-year-old child. To understand brain processing could therefore lead to dramatic improvements in technology. In our group in the Department of Engineering we use theoretical, computational and experimental studies to investigate the computational principles underlying skilled movement in humans (we study undergraduates as a representative sample of human-kind). A major area of our research programme is to understand how the brain deals with the uncertainty inherent in the world and in our own sensory and movement systems. We only know about the world through our senses, but they provide information that is usually corrupted by random fluctuations (known as noise), which lead to variability in our perceptions. For example, if you put one hand under a table and try to localise it on top of the table with your other hand, you can be off by several centimetres. Moreover, when we act on the world through our movement system, the commands we send to our muscles are also corrupted by a variability that leads to movement inaccuracy. Or, as every darts player knows, trying to aim for the same spot over and over again leads to a large spread of the darts. Therefore, this combined sensory and movement variability limits the precision with which we can perceive and act on the world. Remarkably, society rewards those who can reduce their overall variability. If you can reliably hit a small white ball into a hole several hundred yards away using a long metal stick, the financial rewards can be enormous. However, it is not only society that cares about reducing variability. Our work has shown that the brain works hard at reducing the uncertainty and variability in its perceptions and actions. We have shown that our brains implement a branch of mathematics known as Bayesian Decision Theory. The fundamental idea is that you want to generate beliefs about the world – so what are beliefs? Beliefs could be “Am I looking at a cat or a fox?” or “Are my arms in one configuration or another?” And in the Bayesian world we are going to represent beliefs with probabilities – that is, with a number between zero and one, zero meaning “I don’t believe it at all” and one meaning “I am absolutely certain”, with numbers in between giving the grey levels of uncertainty.
The key idea to Bayesian inference is that you have two sources of information from which to generate beliefs. You have data – and in neuroscience, data is what we sense from the world. But there is another source of information – memory – which can give you prior knowledge. You can accumulate such knowledge throughout your life. And the point about Bayesian decision theory is that it gives you the mathematical tools to determine the optimal way to combine your prior knowledge and your sensory inputs to generate new beliefs. An intuitive example will be familiar to the average tennis player. If you want to estimate (generate a belief) about where the approaching ball is going to bounce, Bayes’ rule tells you there are two possible sources of information. There is sensory evidence – you can use visual information and sound information to make an estimate. However, as your senses are not perfect, there will be variability in where you think the ball is going to land, and you can assign a probability to each location. That is the information available on the current shot, but there is another source of information only available from repeated experience in the game of tennis: that is, that the ball doesn’t bounce with equal probability over the entire court during the match. If you are playing against a very good opponent, they may distribute it close to the edge of the court making it hard for you to return. Both these sources carry important information. And what Bayes’ rule says is that you can combine the sensory evidence on the current shot with your prior knowledge to make the optimal estimate of the bounce location. Indeed, in a similar laboratory-based task we were able to show that humans are Bayesian learners, meaning they represent the statistics of the outside world and know about the variability in their own sensory apparatus, allowing them to optimally combine prior knowledge and sensory evidence to generate beliefs.
But beliefs are of no use to us unless they drive actions, so the second part of Bayesian Decision Theory is the decision and subsequent action. There is a problem. Tasks tend to be symbolic – I want to drink, I want to dance – but the movement system has to contract 600 muscles in a particular sequence. In fact, there is a big gap between the task and the movement system that could be bridged in infinitely different ways. For example, even in a simple arm reach there are in fact an infinite number of paths along which I could move my hand. Even if I choose a particular path, I can hold my hand on that path with infinitely different joint configurations. But it turns out that we are extremely stereotypical. We all move in pretty much the same way. So why is it that we move in the particular ways we do? In reality, perhaps we don’t all actually move quite the same way. Maybe there is variation in the population. And maybe those who move better than others have got more chance of getting their children into the next generation. So in evolutionary scales, movements get better. And perhaps in life, movements get better through learning. So what is it about a movement that is good or bad? We have shown that it is the variability in our muscle contractions that is critical to the choice of a particular movement. So when I want my arm to do something, there is a random component that perturbs my movement away from what I desire. However, it turns out that different ways of moving to the same final location engender different amounts of variability. Indeed, a model we developed which proposes we move in a way to minimise the negative consequence of such random fluctuations, thereby minimising variability, was able to explain a wide range of movement patterns. In summary, our brain evolved to control movement. The major intellectual challenge is to understand how we do that. Not only will such an understanding be applicable to robotic technology, but it is also highly relevant for the many diseases which affect movement. By understanding the normal processes that underlie control, we aim for a better understanding of disease processes, leading to better evaluation, treatments and rehabilitation techniques.
You can view Professor Wolpert’s TED talk at tinyurl.com/4xxquv4. For more information go to www.wolpertlab.com CAM 66 37