Prediction mechanisms of the brain: a computational taxonomy

点击翻译按钮获取中文摘要

Our brains receive a continuous torrent of unsorted, noisy data from the senses - changing patterns of light, distributions of sound frequencies, and signals from our skin and muscles about the position and movements of the body. Yet we perceive an ordered and meaningful world of objects, speech and music (not patterns of light and sound energy), in which we produce actions directed towards behavioural goals (rather than outputting jumble of limb locations and muscle tensions). To create order from chaos, the brain is continually constructing and refining models of the sensory world and the animal's or person's actions in it. These models simplify information processing by capturing the meaningful structure in the external world and ignoring information that is unstructured, meaningless or irrelevant. In this sense we can think of brains almost like scientists, actively trying to understand and predict the world around them by simplifying it down to its core elements. In my research, I am trying to answer some rather general questions about how the brain constructs internal models that accurately capture the structure of the external world. I will be targeting three information processing challenges that brains face again and again in different contexts, and trying to work out if the brain has dedicated systems for solving these problems, and if so, how those systems work: - How does the brain focus on relevant variation in stimuli (such as the rate at which cars are approaching a pedestrian crossing) and filter out irrelevant variation (such as the colour of the cars)? - How does the brain optimise processing of relevant stimuli (for example, focussing attention on stimuli where similar stimuli have very different meanings, and so careful processing is required?- How does the brain determine how much to change its expectations about the world, as the environment changes? If I experience a surprising event (such as an increase in traffic on my route to work), how much should I update my expectations for future occasions? Can I voluntarily set my brain to learn faster (for example, if I expect a change in traffic flow because I know that construction work is taking place) and to forget again (when the construction work is over)?My approach to understanding how the brain meets these challenges is to develop algorithms (implemented in computer programmes) that behave in a similar way to human brains in solving these challenges, and then to work out how these algorithms could be computed by populations of neurons in a real brain. Then to test whether the algorithms are correct (that is, whether the brain really works that way), I use my neural network models to generate new predictions about how people will behave in experiments, and how their brain activity will change in different circumstances. I measure brain activity using non-invasive imaging techniques such as MRI. Some of my models also make predictions about how different brain chemicals (neurotransmitters and neuromodulators) affect the processing of information by the brain. To test these hypotheses, I will give healthy volunteers small doses of drugs that affect neurotransmitter concentrations and observe the resulting changes in their behaviour and brain activity.