Biologically inspired framework for learning and abstract representation of attention control

Abstract

The goal of this research is to develop a framework that enables artificial agents to learn active control of their attention as a means toward efficient planning, decision-making, and recognition. The proposed method is inspired by recent findings in psychology and neuroscience that give rise to the assumption that sequential firing of mirror neurons are connected with prediction, recognition, and planning. As these capabilities are connected to active perception, we hypothesize that simulated sequential mirror neurons can provide an abstract representation of learned attention control. The proposed framework consists of three phases. The first phase is designed for learning active control of attention using reinforcement learning. In the second phase, sequential concepts are extracted from the agent's experience and sequential mirror neurons are generated. In the last phase the concepts represented by these sequential mirror neurons are employed for higher level motor-planning and control of attention, as well as recognition. © Springer-Verlag Berlin Heidelberg 2007.