A Biologically Inspired Active Vision Framework for Cognitive Agents

Visual perception facilitates a rich interface to the environment. In this research, we attempt to model it as an active process that organisms engage in, with the goal of operationalizing it in a computational framework. Our motivations lie in the ubiquitous evidence that natural evolution has managed to demonstrate effective solutions to the vision problem. While a growing community of early-vision researchers have started focusing on biologically-inspired approaches, current models do not provide a unifying theory for cognitive control of perception during a task.;An in-depth study of neurophysiological literature reveals a detailed body of knowledge that is as stunning as it is fragmented. We attempt to assimilate these facts into a functionally complete yet abstract model of visual perception. This informs our architectural approach to designing a framework for active vision that subscribes to a strong notion of biological plausibility.;The framework embodies the idea that the key to visual perception in biological systems lies in the complex interconnection of large numbers of simple neuronal processing units. Computational challenges of scale and novel solutions to deal with them are discussed. Applications of the resulting framework help demonstrate its versatility across tasks and its potential to serve as a means of modeling visual phenomena.;Our specific contributions are in the field of attention-based visual perception and cognitive modeling. We demonstrate how bottom-up saliency-driven attentional mechanisms can be modulated by top-down task-specific influences to precipitate efficient visual computation. As part of an interface with high-level systems, we provide a more realistic vision module for cognitive agents that is able to explain characteristics of human performance on vision-dependent tasks.;A tangible outcome of this work is a software bridge that allows a cognitive model to consume raw visual input generated by a standard psychophysical experimentation tool, and interact with it via emulated input devices. This enables modelers to use experiments in the same form as administered on humans, without the additional step of abstracting the interface which may subtly affect results.;Lastly, we identify a number of critical problems that need to be solved in order to make biologically-plausible neuronal architectures more applicable to practical scenarios. We lay down a roadmap for further investigation in this line of work.