| Abstract: | The basic issue of whether mammalian learning in cortex proceedsvia a selection principle, as stressed by Edelman, versus aninstructional one is of major importance. We present here arealization of selection learning in the trion model, whichis based on the Mountcastle columnar organizational principleof cortex. We suggest that mammalian cortex starts out withan a priori connectivity between minicolumns that is highlystructured in time and in space, competing between excitationand inhibition. This provides a "naive" repertoire of spatial-temporalfiring patterns that stimuli and internal pro-cessing map onto.These patterns can be learned with small modifications to theconnectivity strengths determined by a Hobbian learning rule.As various patterns are learned, the repertoire changes somewhatin order to respond property to various stimuli, but the majorityof all possible stimuli still map onto spatial-temporal firingpatterns of the original repertoire. In order to show that theexample presented here is showing true selectivity and is notan artifact of more stimuli evolving into the learned pattern,we develop a selectivity measure. We suggest that some formof instructional learning (in which connectivities are finelytuned) is present for difficult tasks requiring many trials,whereas very rapid learning involves selectional learning. Bothtypes of learning must be considered to understand behavior. |
