396 John R. Platt 



that the slow growth of new cells or random potential connections occurs in 

 parallel, thousands or milHons of cells at a time, while the fast new decisions 

 or perceptions or insights occur sequentially, hooking up one cell at a time or 

 a small group. Such a sequence might resemble the activity-stimulation- 

 proliferation-organization sequence in other tissue. And while this specific 

 suggestion may again be wrong, its accuracy is less important than its general 

 bearing on the time-constant problem, which suggests that epochs of growth 

 may need to be separated from epochs of decision in a biological address- 

 determining network. 



This possibility seems to deserve experimental inquiry. Perhaps our limited 

 time span of intellectual attention, and the 'subconscious' solution of problems, 

 and the role of sleep, especially in the infant, in preparing new cells to be 

 ready for new (waking) connections or learning or decisions, should be re- 

 examined from this point of view. 



C. Artificial Non-addressed Systems 



The truck driver is trained in cliildhood to perceive and respond appro- 

 priately to cars, stop-lights and pedestrians of whatever kind. In this pattern 

 and analogy-perception he excels any arrangement of photocells yet created. 



A pre-addressed decision-net might be able to operate with his small high- 

 way tolerances and high speeds if it had his 10^-element resolving power and 

 wide tield of view. But it would not be safe in the unpredictabihties of the 

 open road. For this job, a non-addressed mosaic is needed, capable of learning 

 new patterns. Otherwise the appearance of a new type of car or a new type 

 of hazard on the road will cause the machine to be sent back to the factory 

 for a complete rewiring of the circuits to establish the new invariances and 

 their analogies with the old cars and the old hazards. 



It is important that the new hazard be recognized by analogy and not by 

 trial and error. Direct highway experience would eliminate quickly a number 

 of types of 'learning' computers that have recently been devised, in which the 

 internal strategies are altered according to experienced successes or failures, 

 but in which there is no pattern-extrapolation or 'insight'. 



The possible construction of artificial non-addressed 10"*- to 10^-element 

 systems with complete decision nets and with 10^- to 10^-element outputs 

 may deserve consideration. Primitive pattern-perceiving networks might be 

 useful for narrowing the band-width of communication channels, if not for 

 crude vehicle guidance. They might be useful internal elements in high-speed 

 analogue and digital computers, where their stupidity could be partly com- 

 pensated by the speed of operation. There they might simphfy the presently 

 elaborate programming operations; and could speed up computations requiring 

 many simultaneous substages of qualitative judgment or identification under 

 distortions or transformations, where the total judgment is more elaborate 

 than can be quickly represented by the coincidence of two digital words. 



The complete theory of artificial non-addressed systems with their many 

 quasi-human characteristics will be fascinating. Evidently in many respects 

 it may be simpler and more physical than present theories of digital computers 

 and single-channel systems. It would include questions of optimization of 

 different aspects of mosaic detection, such as rates and cell sizes, the proper 



