1/8 THE PROBLEM 



had closed the throttle, the same action will now open it; and 

 what was the correct pairing of red and green to push and pull 

 must now be reversed. So the local action in the nervous system 

 can no longer be regarded as ' correct ' or ' incorrect ' in any 

 absolute sense, and the first simple solution breaks down. 



Another example is given by the activity of chewing in so far 

 as it involves the tongue and teeth in movements which must 

 be related so that the teeth do not bite the tongue. No move- 

 ment of the tongue can by itself be regarded as wholly wrong, for 

 a movement which may be wrong when the teeth are just meeting 

 may be right when they are parting and food is to be driven on 

 to their line. Consequently the activities in the neurons which 

 control the movement of the tongue cannot be described as either 

 4 correct ' or * incorrect ': only when these activities are related to 

 those of the neurons which control the jaw movements can a 

 correctness be determined; and this property now belongs, not to 

 either separately, but only to the activity of the two in combination. 



These considerations reveal the main peculiarity of the problem. 

 When the nervous system learns, its behaviour changes for the 

 better. When we consider its various parts, however, we find that 

 the value of one part's behaviour cannot be judged until the 

 behaviour of the other parts is known; and the values of their 

 behaviours cannot be known until the first part's behaviour is 

 known. All the valuations are thus conditional, each depending 

 on the others. Thus there is no criterion for ' better ' that can 

 be given absolutely, i.e. unconditionally. But a neuron must do 

 something. How then do the activities of the neurons become 

 co-ordinated so that the behaviour of the whole becomes better, 

 even though no absolute criterion exists to guide the individual 



neuron 



Exactly the same problem faces the designer of an artificial 

 brain, who wants his mechanical brain to become adaptive in its 

 behaviour. How can he specify the ' correct ' properties for each 

 part if the correctness depends not on the behaviour of each part 

 but on its relations to the other parts ? His problem is to get 

 the parts properly co-ordinated. The brain does this auto- 

 matically. What sort of a machine can be ^Z/-co-ordinating ? 



This is our problem. It will be stated with more precision in 

 S. 1/17. But before this statement is reached, some minor topics 

 must be discussed. 



7 



