Integrating perception and cognition for AGI

Abstract

Current perceptual algorithms are error-prone and require the use of additional ad hoc heuristic methods that detect and recover from these errors. In this paper we explore how existing architectural mechanisms in a high-level cognitive architecture like ACT-R can be used instead of such ad hoc measures. In particular, we describe how implicit learning that results from ACT-R's architectural features of partial matching and blending can be used to recover from errors in object identification, tracking and action prediction. We demonstrate its effectiveness by building a model that can identify and track objects as well as predict their actions in a simple checkpoint scenario. © 2011 Springer-Verlag Berlin Heidelberg.