Inconsistency-induced learning: A step toward perpetual learners

摘要

One of the long-term research questions in machine learning is how to build never-ending learners. The state-of-the-practice in the field of machine learning thus far is still dominated by the one-time learner paradigm: some learning algorithms are utilized on data sets to produce certain results, and then the learner is put away and the results are put to work. Such a learn-once-apply-next (or LOAN) approach may not be adequate in dealing with many real world problems and is in sharp contrast with human's life-long learning process. On the other hand, learning is often brought on through some stimulus. In this paper, we describe a framework for inconsistency-induced learning. The framework relies on utilizing inconsistency as learning stimulus and inconsistency resolution as impetus for continuous learning. The framework hinges on recognizing inconsistency in information or knowledge, identifying the cause of inconsistency, revising beliefs to explain, resolve, or accommodate inconsistency. The perpetual learning process is triggered by an agent encountering some antagonistic circumstance, and is embodied in the continuous inconsistency-induced belief revisions. Though there can be other stimuli to learning, we believe that inconsistency-induced learning can be an important step toward building perpetual learning agents.