Biologically Inspired Task Abstraction and Generalization Models of Working Memory

摘要

We first present a model of working memory that affords generalization. By separating stimuli in such a way that filler representations may flow through the model based on the state of gates, which are opened or closed in response to role signals, an action selection network is afforded the ability to learn a response to fillers that is independent of the roles in which they were encountered. Next, we present n-task learning, an extension of temporal difference learning that allows for the formation of multiple policies based around a common set of sensory inputs. In order to allow for state inputs to take on multiple values, they are joined with an arbitrary input called an abstract task representation. Task performance is shown to converge to optimal for a dynamic categorization problem in which input features are identical across all tasks.