Learning to perceive through action : computer vision in the autonomous empirical reasoning architecture

Abstract

This research explores the topic of the autonomous active percetual learning in an artificial agent. Cognitive architectures that can support such learning are few—one of them is the Autonomus Empirical Reasoning Architecture (AERA), which uses causal learning and reasoning to build knowledge on its own, from a seed. A new framework has been implemented for enabling AERA to cumulatively learn to see from experience by modeling how its actions affect perceptual information to produce novel and useful knowledge. We propose a vision system that bootstraps perceptual learning solely from image segmentation and feature detection. Since cumulative situated learning must operate in real-time and be able to adapt to diverse environments, the system we have developed is both fast enough for real-time operation and sufficiently general to align with this requirement. During learning, the outputs of these two visual sensory sources are combined and transformed, autonomously, into high-level causal representations suitable for the AERA architecture to use for practical action. To evaluate our approach, we conducted two proof-of-concept experiments on individual components of the system and scenarios that assess the whole vision system and its integration with AERA. Our findings show that this is a viable method for integrating visual perception into the AERA architecture.