Neural associative memory with optimal bayesian learning
Neural Computation
| Hi-index | 0.00 |
Hebbian synapses lie at the heart of most associative matrixmemories (Kohonen 1987; Hinton and Anderson 1981) and are alsobiologically plausible (Brown et al. 1990; Baudry and Davis1991). Their analytical and computational tractability make thesememories the best understood form of distributed informationstorage. A variety of Hebbian algorithms for estimating thecovariance between input and output patterns has been proposed.This note points out that one class of these involves stochasticestimation of the covariance, shows that the signal-to-noise ratiosof the rules are governed by the variances of their estimates, andconsiders some parallels in reinforcement learning.