Abstract :
The field of computer vision, in general, is focused on achieving maximal computational efficiency with little priority towards mimicking biological vision. Biologically grounded vision algorithms such as Poggio´s layered features and Deep Learning Networks, on the other hand, are geared towards achieving recognition performance and predictable artefacts similar to the human vision system. Nevertheless, there is a dearth of computer vision algorithms and associated robotic systems that incorporate human-like spatio-temporal behaviors in visual processing, which has been conveniently modeled in symbolic systems such as ACT-R. In this paper, we explore the possibility of grounding a recently developed semantic cognitive vision theory - the k-TR theory of affordances for visual perception, using biological models of visual cue activation, thus attempting a marriage between one family of computer vision processing algorithms and ACT-R based spatio-temporal behavior models. Various biological effects such as frequency and recency of feature activation, cognitive concept linkages and spreading activation, partial matching of concepts and visual features, effect of noise at various levels of perception and cognitive analysis, are taken into account. While it can be seen that the performance of such a cognitive k-TR system grounded in biological/brain activation models is inferior in recall performance to an infinite memory, infinite time computational k-TR system, the temporal characteristics in terms of recall times and memory management of the biologically grounded system are superior to the purely computational model when deployed in a semantic context. The implementation of such a system is thus an important step towards the development of human-like behavior and performance in cognitive robots.
