Neurocomputational Mechanisms for Generalization During the Sequential Learning of Multiple Tasks.
Ashish Gupta, David C. Noelle · 2005
Abstract. Traditional artificial neural network models of learning suffer from catastrophic interference. They are commonly trained to perform only one specific task, and when trained on a new task, they forget the original task completely. It has been shown that the foundational neurocomputational principles embodied by the Leabra cognitive modeling framework are sufficient to overcome this limitation significantly. Leabra is consistent with known biological properties of the neural networks of the brain. In particular, this framework includes both fast lateral inhibition and a local synaptic plasticity model that incorporates both correlational and error-based dynamics. It has been shown that the use of sparse internal representation significantly reduces the problem of catastrophic interference during sequential learning of multiple motor skills. We have also provided prelimilary evidence that Leabra is able to generalize the sub-sequences of motor skills, when doing so is appropriate. In this paper, we provide a detailed analysis of the extent of generalization possible with Leabra during sequential learning of multiple tasks. For comparison, we measure the generalization shown by a backpropagation of error learning algorithm based artificial neural network.