Minimally disturbing learning

Abstract

We have undertaken the study of methods for avoiding catastrophic forgetting in feedforward neural networks, without sacrifying the benefits of distributed representations. We formalize the problem as the minimization of the error over the previously learned input-output (i–o) patterns, subject to the constraint of perfect encoding of the new pattern. Then we transform this constrained optimization problem into an unconstrained one. This new formulation naturally leads to an algorithm for solving the problem, wihch we call Minimally Disturbing Learning (MDL). Some experimental comparisons of the performance of MDL with back-propagation are provided which, besides showing the advantages of using MDL, reveal the dependence of forgetting on the learning rate in back-propagation.