Perpetual Learning through Overcoming Inconsistencies

This paper provides a panoramic view on perpetual learning through overcoming inconsistencies. The approach elevates learning stimuli to the first class status and considers inconsistencies as a special type of learning stimuli. As telltales, inconsistencies indicate that an agent is operating at its knowledge boundaries, necessitating subsequent learning episodes. Learning amounts to finding ways to circumvent inconsistencies. We describe a framework called inconsistency induced learning, or i²Learning, and discuss several specific learning algorithms for it. The perpetual nature is embodied in the fact that i²Learning accommodates the open-ended sequence of learning episodes.