CosynePoster5

Poster I-1 at Cosyne 2018

My approach to Organic Learning will be presented in a poster at Cosyne 2018

Look for it at location I-1 in the March 1st Poster Session

Abstract

In this modeling study, simple patterns of local connections between neurons embedded in topographically arranged layers provide built-in object recognition capability to a spiking neural network.  The strategy is inspired by recent detailed examinations of how connections between specific cell types and at specific dendritic locations implement computational functions in motion detection.  For object recognition, straightforward structures create a sparse population code of topographically organized feature detectors.  The main result of this study is that having such a structure allows a spiking neural network to learn to recognize objects rapidly and without assuming backpropagation of information in the network.   Three dimensional structures that implement specific response patterns are illustrated, showing that for topographically organized neurons, certain patterns of connections are equivalent to computation. The resulting population code is easy to interpret, like the responses of visual neurons recorded in vivo, and the network is endowed with responsiveness to appropriate stimuli from inception rather than being initially random in its responses.  Because of these characteristics, such structures may serve as means of enabling object recognition in visual systems with most of the necessary information transferred in the genome, and only limited task specific learning after an individual organism is born.  This breaks with the standard of statistical learning theory but reflects the observation that most organic brains must compete to survive immediately upon birth, with little chance to learn. A secondary topic of the study is that the network design relies on the equivalence between certain classes of binary and spiking neural network models, and the biophysical limits on this equivalence is explored.  This method of analyzing spiking neural networks may be applicable to a broad class of networks.