COMPARING THE BRAllSr WITH MACHINES — ^MacKAT 235 



Among the less explicit, thougli nonetheless important, restrictions 

 on modelmaking is the requirement that the brain as modeled must 

 have been capable of growing that way. This has led some of us to in- 

 vestigate the possibilities of self-organizing statistical, or "proba- 

 bilistic," models. By a statistical model may be meant («) one whose 

 elements function deterministically but whose detailed interconnec- 

 tions are assumed to form more or less randomly (Ashby, 1952), or 

 (h) one whose elements function indeterministically, to some control- 

 lable extent (MacKay, 1952). 



The model that promises to be most useful is statistical in both these 

 senses (Wisdom et al., 1952; MacKay, 1952). Its basic feature is the 

 ability to adjust the rules of its own activity according to the degree 

 of success it attains; but it does this, not by throwing switches, but 

 by adjusting the relative probabilities of dili'erent patterns of activity. 

 Its rules, in other words, are grown statistically as a result of its own 

 trials and errors. The interconnections between elements are not hard 

 and fast, but each link has an adjustable probability of functioning, 

 which will in general depend on the location and the timing of all 

 neighboring activity. 



You might picture a typical element in such a model as something 

 like a gun with a rather uncertain trigger. The farther the trigger is 

 pulled, the greater the probability that the gun will go off. If we sup- 

 pose that when a gun goes off, its firing alters the tension in the trigger 

 of another gun, without necessarily firing it, we have the basic model 

 of a mechanism that can alter the probabilities of its own activities. 



I wish it were possible to go into more detail. But perhaps you can 

 see roughly how a network of such elements, if supplied automatically 

 with signals to indicate success or failure, could use them to alter prob- 

 abilities so as to grope its way more and more quickly into a pattern 

 of activity to "match" any incoming patterns that persisted in recur- 

 ring. Its good guesses, so to speak, could be made to persist, and its 

 bad ones to drop out. But because its activity would never be unique- 

 ly determined by the input, there would always be room for occasional 

 spontaneous attempts to try something new as a "matching response" 

 to the flux of events. With a hierarchical structure, these attempts 

 could include guesses about the relative probabilities of other guesses, 

 so that abstractions of any order could be evolved (Wisdom et al., 

 1952). 



Such a model is sufficiently general to be capable of any activity we 

 may care to specify, and it bears promising resemblances to the human 

 nervous-and-humoral system. But if it is fortunate enough to survive 

 the next 20 years, our only reasonable certainty is that the model will 

 have grown almost out of recognition. If it has not, it will have been 

 of small service. 



