Some Statistical Results for Winner-Take-All Network with Sparse RandomInnervation and Coactivity-Based Learning

摘要

Some basic statistical features of the winner take all network with sparse randominnervation are examined, and the effects of a simple coactivity based learning rule discussed. The results are applicable to a range of neural network models employing sparse random connectivity, coactivity based learning, and winner take all networks as part of their architectures. Results suggest constraints on network size and other parameters if a network is to participate in a clustering operation, as in the Granger, Lynch, and Ambros Ingerson (GLA) model of early olfactory processing. It is shown how learning can result in pattern capture, and two mechanisms of capture are identified: capture by precedence, in which a pattern is captured from one cell by another by virtue of precedence in the order of pattern presentation during training, and capture by overtake, in which input to the capturing cell for the pattern to be captured grows faster than that for the original winner, by virtue of the growth of synaptic weights from input fibers also active on other patterns which elicit wins from the capturing cell.