Can grammar and complexity bridge the gap between psychology and biologically plausible neural networks?

There are two approaches to neural network modelling in biology. One can take an existing relatively well understood artificial neural network and determine how its behaviour is changed if it is made more consistent with the known anatomy and physiology of the brain. On the other hand, one can start from known biology and consider what types of useful behaviour such networks might produce. However both of these approaches have drawbacks. The author argues that these difficulties can be ameliorated if we characterise the behaviour of models in terms of the class of grammar and the complexity of the simplest descriptions which can be constructed of their dynamics. The author's approach describes the types of functions that networks of a given type might be capable of in general, rather than identifying the implementation of a particular function in a particular network. Comparison of this type of characterisation with a similar characterisation of algorithmic descriptions of the functions which we expect areas of the brain to perform can act both as tests of functional hypotheses in psychology and of models in computational neuroscience.<
>