Before I get started, I’d just like to quickly point to a video piece we managed to get into New Scientist! Please see here for this very exciting media coverage!
Now, I wanted to write a little bit about how Inside TRAK has been useful for demonstrating why computational neuroscience is important. You can think of computational neuroscience as a field in which we are concerned with the information processing capabilities of the brain. We aim to understand behaviour in terms of the chemical and electrical activity of different components of the central nervous system (brain and spinal cord). This doesn’t just mean associating certain behaviours with particular regions of the brain, though this is an important part of neuroscience and often helps guide computational work. Rather, we look at the manner in which different regions of the brain react to different types of information, where and how this information is further propagated, and we attempt to understand how the processing that we observe might be responsible for the behaviours we observe. This can be contrasted with psychology, for example, which tends to use a more abstract level of explanation. So, whereas a psychologist may be more concerned with explanations of behaviour in terms of constructs like attitudes and emotions, a computational neuroscientist may prefer to interpret those same behaviours in terms of the efficiency with which certain neurons propagate information in different contexts. It is important to realise however, that these approaches are complimentary, and can help guide each other towards better and more comprehensive theories of why we behave the way we do.
In computational neuroscience, sometimes we focus on the significance of the activity of individual neurons; other times on that of the average activity in larger structures composed of millions of neurons. The latter is referred to as the ‘systems level’, and is the level with which Inside TRAK is concerned. Comp neuro research involves creating models of these brain systems and using them to simulate the brain activity that we observe in the biological system. This clearly requires a detailed understanding of the neural systems in question. Using brain imaging and other techniques, experimental neuroscientists have gathered mountains of data revealing which parts of the brain are active when we perform particular tasks. Simulations that reproduce this neural activity – like the one powering Inside TRAK – help us to test theories of how our brains work. Inside TRAK is a great example of this. As the model that’s powering it is based on the real neural systems we know to be involved in deciding where we look, if the model can’t drive the robot to look around, we know that we’re interpreting something wrong. Perhaps it’s just a trivial problem; perhaps our model is too high level and we haven’t modelled sufficient detail. Alternatively, we may have completely misunderstood something about the way the system works.
However, this is by no means the only reason why it’s important. It’s not until you try to emulate a behaviour that you are forced to address the assumptions and omissions you might have made in your theory. I like to call it the ‘magic box’ problem. It’s all too easy for us to brush over what’s really happening in the brain by explaining behaviour in terms of things we all instinctively think we understand, but find it very hard to explain what we really mean. If we think about this in terms of Inside TRAK, we might say that the robot weighs up the pros and cons of looking left or right, and that’s how it decides to look one way or the other. We may feel as though we’ve explained what’s going on as ‘weigh-up-pros-and-cons’ is a very instinctive process to us. This, however, is the ‘magic box’. Several options are introduced to the magic box, something happens inside it, and out pops the decision. Until we can simulate what’s happening in that box in terms of the brain though, we don’t understand what the brain is doing. Observing the behaviour of models we have built also helps us to generate predictions about what is happening in the brain, as we can poke and prod it in ways we can’t do with biological systems. We can systematically vary parameters that are representative of very real things in the brain, and examine the subsequent behavioural effects in order to further understand the biological systems on which our models are based. I could go on, but this post is probably dragging as it is, so I’ll stop there and point you to an excellent essay by Joshua Epstein about why modelling is important. It’s neither long nor boring (possibly unlike this post), so I really recommend it.
Next post: a more detailed look at the function of iTRAK with some pretty pictures and everything.