WHAT IS THE CATCH IN NEURAL NETS?
DOI link for WHAT IS THE CATCH IN NEURAL NETS?
WHAT IS THE CATCH IN NEURAL NETS? book
In the mind-brain sciences we often think of brains as being supremely complicated machines, seen in recent years as super-computers. So, in a sense, brain science is a branch of engineering. I expressed this over thirty years ago in a paper called 'The Brain as an Engineering Problem' (Gregory 1961). A central point was how difficult - impossible - it is to localize brain function, or say where processes are going on, without 'engineering' knowlege of the internal functions. This kind of description overcomes the Phrenologist's Fallacy of supposing that internal functions (in 'bumps') are directly related to external functions of behaviour. We may say without any problem that components in a brain or a machine (nerve cells, or transistors) are spatially localized; but it is impossible to say, for example of an engine, where its power is localized as many component parts contribute to the final result. Brain functions are extended over many interacting neural components, which makes it hard to say where intelligence, or perception, or whatever, lie in the brain. This can make interpreting brain ablation studies very difficult. To quote the example I used in another paper (Gregory 1958):
If any changes take place upon removal of part of the brain, the changes are either (a) loss of some feature of behaviour, or diminution or worsening of some skills, or (b) introduction of some new behavioural features. Now it is often argued that if some part of behaviour is lost . . . the causal mechanism necessary for this behaviour is localized in the affected region. But does this follow? . .. We can take an example from radio engineering.