ABSTRACT
Long-term potentiation (LTP), the current leading candidate biological mechanism for learning, has as physiological parameters a step size (the amount by which a single potentiation episode increases synaptic efficacy) and a ceiling (the level at which further potentiation results in no further increase). Recent findings have identified endogenous agents that modulate these two parameters of LTP 1 . A statistical model of a cortical network was used to predict the effect these parameters have on categorizing input patterns. Results show that a larger step size (i) leads to faster learning and (ii) causes the network to form narrower (more restrictive) categories. Furthermore, results show that a larger ceiling (i) leads to slower learning and (ii) causes the network to form more general (less restrictive) categories. The effects of step size and ceiling interact, and their relationship to category breadth is highly nonlinear. Settings of the ratio of step size to ceiling are identified that optimize the tradeoff between generalization and learning rate.
