Combining Hebbian and reinforcement learning in a minibrain model

TL;DR

This paper presents a toy neural network model combining Hebbian and reinforcement learning to address path interference, showing optimal learning performance when the Hebbian term is balanced with the reinforcement term.

Contribution

It introduces a combined Hebbian-reinforcement learning rule and identifies the optimal ratio for reducing learning steps in a neural network model.

Findings

Optimal ratio of Hebbian to reinforcement learning reduces learning steps

Adding Hebbian term mitigates path interference

Balance between Hebbian and reinforcement terms is crucial

Abstract

A toy model of a neural network in which both Hebbian learning and reinforcement learning occur is studied. The problem of `path interference', which makes that the neural net quickly forgets previously learned input-output relations is tackled by adding a Hebbian term (proportional to the learning rate $\eta$) to the reinforcement term (proportional to $\rho$) in the learning rule. It is shown that the number of learning steps is reduced considerably if $1/4 < \eta/\rho < 1/2$, i.e., if the Hebbian term is neither too small nor too large compared to the reinforcement term.