|
|
|
|
|
|
|
performance. Just as in games and searches, there is the problem that individual responses do not often directly yield need satisfaction. However, the internal model discussed in connection with property (6) enables the CNS to constantly improve performance in the absence of current need satisfaction. Two kinds of improvement are possible. First of all, cell assemblies typically respond to too broad a range of situations when first formed, yielding inconsistencies in the model. That is, situations activating the same combinations of cell assemblies, and hence the same responses, are followed by radically different outcomes. The remedy here is much like that for inadequacy of detectors discussed in the illustration on searches. Because of the inconsistencies new associations are formed between the cell assemblies involved, causing them to split and recombine so that their responses are more discriminative. (Hebb calls the related procedures fractionation and recruitment.) The second kind of improvement consists in "filling in" the modelgenerally there will be many situations where no expectations or predictions have been developed. This clearly provides an important role for curiosity. The CNS must experience a wide enough range of situations to provide an adequate repertory of relevant temporal sequences. Just as with the coadapted sets of genetics, the basic laws of cell assemblies permit flexible recombination (association) under environmental (sensory) guidance, the actual combinations being influenced by the parts of the model (associations) already extant. In this way a tremendous range of useful procedures can be formed from relatively few elements. More importantly, a single experience then constitutes a trial of a great many relevant associations, just as in genetics a single organism tests a great many coadapted sets. Property (3) assures that many associations will be tested and modified. The ultimate "survival" of various combinations of assemblies is determined by their consistency within the model and their success in contributing to learned or unlearned need satisfaction. |
|
|
|
|
|
|
|
|
While the foregoing analogies are ready offshoots of the formal framework, the basic task of theory in this area is quite difficult. It must enable one to judge whether proposed mechanisms for CNS operation permit the learning rates, utilization of cues, transfer of learning, etc., that one actually observes. How does the CNS maintain its rapidity and appropriateness of response, while extending its breadth and filling in its model of the environment? Section 8.4 indicates one way in which concepts from the  framework can be brought to bear. In particular, the robustness of reproductive plans, when interpreted in this area, indicates some promising directions, but we even lack good general measures of performance here. A kind of error function based upon average need levels might be interesting for organisms not quite so efficient as man at keeping their primitive needs satisfied. A criterion  could then be formulated, much as it was in the optimal control |
|
|
|
|
|