Neuron Selectivity as a Biologically Plausible Alternative to Backpropagation

In this paper, we aim to develop alternative methods to backpropagation that more closely resemble biological computation. While backpropagation has been an extremely valuable tool in machine learning applications, there is no evidence that neurons can back propagate errors. We propose two methods intended to model the intrinsic selectivity of biological neurons to certain features. Both methods use selectivity matrices to calculate error and to update synaptic weights in different ways, either by comparing neuronal firing rate to a threshold firing rate algorithm or computing error from a generated score scoring algorithm. We trained and tested networks that used either of the two algorithms with the MNIST database. We compared their performance with a multilayer perceptron network with 3 hidden layers that updated synaptic weights through typical error backpropagation. The backpropagation algorithm had a test error of 1.7%. Networks that updated synaptic weights based on the scoring method gave a test error of 2.0%, and networks using the firing rate method 4.3%. We were thus able to develop more biologically plausible models of neural networks in the brain while obtaining performance comparable to typical backpropagation algorithms.