The vision of the brain as a machine for, cognitively speaking, finding correlations in its inputs so as to be able to create entities from the sensorial chaos has its attractions. We are, it must be confessed, talking about what has been called concept formation here, although the term must fill any mathematician, or even anyone with a decent feel for language, with fury and loathing. The point of `extracting features', to use another abominable contortion of the inarticulate, in this way, is of course to do more than classify, it is to predict, contingently; it is to survive. It would seem however, that the process of extracting and symbolising entities via layers of neurons has to be a fairly basic piece of machinery with which to accomplish survival in this present Universe.
In this chapter I indicated initially the drawbacks of the kind of static recognition I had used at the start of the book, and which are the commonplace methods of the times. I then drew attention to the inspiration of King Sun Fu that there was some similarity between the structure of natural language and that of certain images, and pointed out certain drawbacks in his attempt to provide algorithms for using this insight to understand images. I assured the reader that there was a formal theory couched in Category Theoretic terminology which allowed one to exploit the structures Fu had observed, and I undertook to convey the elements of this theory by practical exampole rather than formal mathematics. The remainder of the chapter consisted of my trying to fulfil that promise, with what success the reader is in a position to pass judgement, and in a final section I indicated that there was a link between the statistical methods of mixture modelling and what neurons might do in the Central Nervous System - in other words, paradoxically, statistical pattern recognition may be closer to what neurons do than neural net pattern recognition. I then suggested that there was a link between the UpWrite processes and the layering of neurons in the Nervous Systems of the higher animals.
The theory of cognitive processes and their implementation which this implies, may be tested empirically by building syntactic pattern recognition systems. There is some prospect of some of the work which has applied these ideas, actually making a few bucks for the present writer in the not too distant future. Very, very few bucks are likely to actually get past the tax man and his emulators. Of course, they would only be spent on Havana cigars, or even less respectable forms of self-indulgence, so perhaps it is no bad thing that the customers are all close to broke, or claim to be.