394 John R. Platt 



network, where the inputs do not permanently modify the connections. The 

 latter must have a separate storage unit capable of modification and isolated 

 from the main network except for controlled temporary interactions. So our 

 electronic computers have their fast-access and slow-access storage units. 

 Our social decision networks have their files and libraries more or less insulated 

 from the functioning decision-personnel. This necessary but misleading 

 subdivision of our artificial systems into operating units and storage units 

 may be the reason why so many investigators in the past have searched — 

 unsuccessfully — for a special memory organ in the human brain. Functional 

 storage may be a typical property of biological systems, a further manifestation 

 of their usual simplicity and efficiency. 



As an example, consider the genetic material of the cell, which at the present 

 time is supposed to consist of a few species-specific macromolecules, such as 

 DNA or RNA. In a newly-formed cell, such a molecule has two functions 

 (although they might not be separate functions): to initiate the steps up the 

 ladder of chemical syntheses of specific cell materials; and to duphcate itself. 

 But this is functional storage: the chemical structure and reactions are the 

 expression of the heredity; the heredity is the chemical structure. 



Likewise in the production of antibodies by antigens, the chemical record 

 of the first antigenic experience is preserved in the antibodies (or in the chemical 

 information in the antibody-producing cells), ready to find instant chemical 

 expression when a second essentially identical experience occurs. The record 

 is the specific chemical protection; the protection is the record. 



On a grosser scale, evolution is functional storage. The coming of the 

 cold is shown in our fur and feathers and families. The record of the ancient 

 temperatures and salinities may be in our blood and tears. 



The speed and efficiency of social decision networks might be increased 

 if they could incorporate this lesson, and replace some of their file cabinets 

 by continuously repeated appropriate functional modification in the decision 

 channels. 



Time Constants 



Address-determination must go on at a certain regulated rate. This is 

 probably faster for early-stage neurons and slower for later ones, but the order 

 of magnitude should be well-defined for a given network. 



In the adult human brain, the indications are that roughly 50 milhseconds 

 elapse between distinguishable perceptions or decisions — one 'moment', in 

 the Stroud terminology. This is of the order of fifty of the milHsecond repetition 

 intervals or synapse intervals of an individual cell, which seems to be a reasonable 

 relationship (1). Knowing this time constant, we can make some numerical 

 estimates of brain rates and capacities. These estimates are naive and probably 

 false in detail, but they are explicit and rather instructive. 



Thus suppose that there is one new perception every moment and that 

 it may be preserved in a memory, represented by a single changed neural 

 connection. The now-classical experiment of micro-electrode stimulation 

 during brain surgery shows at least that if certain points are stimulated, a 

 complete, detailed and specific memory is indeed evoked. Combining this 

 with the working hypothesis suggested by Quastler and others (18), that the 



