392 John R. Platt 



may also be a necessary normal and even rather simple aspect of the neural 

 connection process in any system that can determine its addresses. 



B. Growth Characteristics 

 Learning 



We saw earher that a non-addressed system must be initially incompetent 

 and needs a long learning time. This learning or address-determination requires 

 inputs containing pattern regularities — that is, experience. For the retina, 

 the experience may be generated by the external environment alone, or from 

 this environment as scanned by the eyeball; in either case it is external to the 

 retina. 



In either case it generates spaces and metrics independent of the retina. 

 Scanning is probably the visual counterpart of exploratory oral and manual 

 manipulation which defines the 'spaces' of taste and touch. Probably the 

 'externahty' of the visual metric, plus the simplicity and universal identity 

 of the scanning operations of all eyeball-spheres about their centers, help to 

 account both for the Kantian a priori character, and for the pubhc and universal 

 character, of visual space. This is contrasted with the situation, for example, 

 in vocal or tone-quality space, which depends on the complex interaction of 

 hidden muscular movements, and is perhaps the most incommunicable of 

 our public spaces. 



The network can learn only those types of regularities it has experienced. 

 Two networks should develop somewhat different pattern perceptions if their 

 environmental regularities or scanning schemes are systematically altered. 

 A non-addressed network which is forced to operate for a long time in a 

 structureless environment, like a bhndfolded and insulated animal or human 

 in the Riesen and Hebb experiments, should and does have seriously defective 

 pattern-perception and response. One can see how the formation of simple 

 and accurate early-stage addresses in a network would be very important 

 in facilitating fast accurate pattern-perception at later stages. 



This picture of non-addressed learning exemphfies Hebb's conclusion (17) 

 that many adult pattern-perceptions having introspectively the most instinctive 

 and self-evident or necessary character are in fact perceptions that had to be 

 learned at some very early age. It is early experience that selects the address- 

 connections that are to be permanent; it is the permanent address-connections 

 that create expectations and pattern-organizations in later experience. 



Nevertheless, there is a double paradox in the present picture, (a) There 

 are possible external input experiences that cannot determine address connections. 

 And (b) There are internal pattern-perceptions, just as there are eyeball-shapes, 

 which have grown or have been learned and yet have not been determined by 

 the particular experiences or motions that contributed to the learning process. 



Both these conclusions would follow from the operator equations of the 

 last section that were supposed to represent network structure. The first 

 point is obvious and almost trivial. External field patterns and their images 

 are usable only if they fall within the limitations set by the network growth 

 mechanisms and assembly principles. Motions of too high a velocity, patterns 

 of far too coarse or far too fine a structure, fields of diffuse clouds with no 



