The Role of Neurocomputational Principles in Skill Savings

Humans exhibit savings in skills. A skill is rarely forgotten completely, even if it remains unused for long periods. Also, the reacquisition of a skill to its previous level of competence is faster than initial skill learning. Traditional artificial neural network models of skill learning have been unable to exhibit savings comparable to that seen in humans because they 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. A number of specialized connectionist architectures and learning rules have been suggested as means to avoid catastrophic interference. Instead of introducing such a new mechanism, we have investigated the degree to which the foundational neurocomputational principles embodied by the Leabra cognitive modeling framework are sufficient to ameliorate catastrophic interference. In particular, this framework includes both fast lateral inhibition and a local synaptic plasticity model that incorporates both Hebbian and error-based dynamics, grounded in known properties of cortical circuits. In this paper, we provide evidence that these fundamental computational properties of neural circuits can support savings during sequential learning of multiple motor skills.