Re-framing the characteristics of concepts and their relation to learning and cognition in artificial agents

• Published in: Cognitive systems research Vol. 44; pp. 50 - 68
• Main Authors: Olier, Juan Sebastian, Barakova, Emilia, Regazzoni, Carlo, Rauterberg, Matthias
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2017
• Subjects: Artificial intelligence; Cognition; Concepts; Conceptual representations; Machine Learning; Robotics; GM; AI; DL; Cognition; VA; Machine Learning; Robotics; CS; Concepts; RNN; VRNN; Conceptual representations; GC; Artificial intelligence; FC; SGP; EC
• ISSN: 1389-0417; 1389-0417
• DOI: 10.1016/j.cogsys.2017.03.005

In this work, the problems of knowledge acquisition and information processing are explored in relation to the definitions of concepts and conceptual processing, and their implications for artificial agents. The discussion focuses on views of cognition as a dynamic property in which the world is actively represented in grounded mental states which only have meaning in the action context. Reasoning is understood as an emerging property consequence of actions-environment couplings achieved through experience, and concepts as situated and dynamic phenomena enabling behaviours. Re-framing the characteristics of concepts is considered crucial to overcoming settled beliefs and reinterpreting new understandings in artificial systems. The first part presents a review of concepts from cognitive sciences. Support is found for views on grounded and embodied cognition, describing concepts as dynamic, flexible, context-dependent, and distributedly coded. That is argued to contrast with many technical implementations assuming concepts as categories, whilst explains limitations when grounding amodal symbols, or in unifying learning, perception and reasoning. The characteristics of concepts are linked to methods of active inference, self-organization, and deep learning to address challenges posed and to reinterpret emerging techniques. In a second part, an architecture based on deep generative models is presented to illustrate arguments elaborated. It is evaluated in a navigation task, showing that sufficient representations are created regarding situated behaviours with no semantics imposed on data. Moreover, adequate behaviours are achieved through a dynamic integration of perception and action in a single representational domain and process.