Hebbian Inspecificity in the Oja Model

TL;DR

This paper extends the classical Oja model to include Hebbian inspecificity, analyzing how crosstalk affects convergence to principal components in neural learning, with implications for biological plausibility.

Contribution

It introduces an error matrix to model Hebbian inspecificity in the Oja model and analyzes its impact on learning outcomes and convergence.

Findings

Learning becomes less effective with increased inspecificity.

Useful learning persists up to a certain level of crosstalk.

The angle between PC1 and the learned eigenvector increases with inspecificity.

Abstract

Recent work on Long Term Potentiation in brain slices shows that Hebb's rule is not completely synapse-specific, probably due to intersynapse diffusion of calcium or other factors. We extend the classical Oja unsupervised model of learning by a single linear neuron to include Hebbian inspecificity, by introducing an error matrix E, which expresses possible crosstalk between updating at different connections. We show the modified algorithm converges to the leading eigenvector of the matrix EC, where C is the input covariance matrix. When there is no inspecificity, this gives the classical result of convergence to the first principal component of the input distribution (PC1). We then study the outcome of learning using different versions of E. In the most biologically plausible case, arising when there are no intrinsically privileged connections, E has diagonal elements Q and off- diagonal elements (1-Q)/(n-1), where Q, the quality, is expected to decrease with the number of inputs n. We analyze this error-onto-all case in detail, for both uncorrelated and correlated inputs. We study the dependence of the angle theta between PC1 and the leading eigenvector of EC on b, n and the amount of input activity or correlation. (We do this analytically and using Matlab calculations.) We find that theta increases (learning becomes gradually less useful) with increases in b, particularly for intermediate (i.e. biologically-realistic) correlation strength, although some useful learning always occurs up to the trivial limit Q = 1/n. We discuss the relation of our results to Hebbian unsupervised learning in the brain.