Generalization and interference in human motor control

Abstract

Humans (and robots) interacting with the environment have to deal with a continuous stream of sensory inputs in an incremental fashion. Such systems face two fundamental issues: (1) they must acquire new skills in a cumulative fashion, that is exploiting previous knowledge to learn new behaviors, and (2) they must avoid the so-called catastrophic interference, where learning new knowledge destroys existing memories. Here, we analyze the problem from the perspective of biological motor control. We first review experimental results of consolidation of procedural memories and the factors affecting it. Then the problem of generalization and interference is examined together with some interpretations in terms of computational models. Finally, we present some possible approaches to the issue of learning multiple tasks while avoiding catastrophic interference in bio-inspired learning architectures.