Concept-Centered Knowledge Representation

A long-standing paradigmatic debate in artificial intelligence is that between the so-called ‘symbolic’ and ‘connectionist’ approaches to knowledge representation. Both approaches propose principles for how general, domain-independent knowledge might be handled in autonomous agents; however, both have limitations, and in most respects, what one brings to the table is sufficiently incompatible with the other to prevent joining their best properties. We propose a third approach of knowledge representation that removes these conflicts and limitations and offers a new path for artificial general intelligence (AGI). Centered around a framework of ‘adaptive concepts’ that uniformly capture numerous kinds of knowledge in latent dynamic conceptual graphs, our concept-centered knowledge representation (CCKR) meets fundamental assumptions that prior approaches do not address and that we argue are necessary for general intelligence. Here we explain CCKR and present arguments for these claims, resting in part on the results from two implemented experimental AGI-oriented systems, the Non-Axiomatic Reasoning System (NARS) and the Autocatalytic Endogenous Reflective Architecture (AERA)