Representation Problems in Machine Learnin: A Proposal

THE PROJECT INVESTIGATES METHODS FOR AUTOMATICALLY CONSTRUCTING OR SHIFTING REPRESENTATIONS FOR MACHINE LEARNING. THE REPRESENTATION IN WHICH GENERALIZATIONS MUST BE EXPRESSED IS A FUNDAMENTAL BIAS THAT STRONGLY INFLUENCES THE INFERENCES THAT A LEARNING PROGRAM MAKES. FOR EFFECTIVE LEARNING, A PROGRAM NEEDS AN APPROPRIATE REPRESENTATION. CURRENTLY, LEARN- ING PROGRAMS ARE NOT ABLE TO FIND AN APPROPRIATE REPRESENTATION. INSTEAD, A PERSON MUST HANDCRAFT A REPRESENTATION AND THEN ASSESS THE EFFECTIVENESS OF THE LEARNING FOR THAT. AN INTELLIGENT LEARNING PROGRAM SHOULD BE ABLE TO PERFORM THIS ACTIVITY AUTONOMOUSLY. ISSUES ADDRESSED INCLUDE USE OF MULTIPLE REPRESENTATIONS, CONCEPT FORMA- TION, FEATURE GENERATION, FEATURE DISCOVER, REPRESENTATION EVALUATION, AND EVALUATION FUNCTION LEARNING. EXPECTED RESULTS INCLUDE IMPROVED ABILITY TO CONSTRUCT HYBRID REPRESENTATIONS, TO CONSTRUCT NONLINEAR EVALUATION FUNC- TIONS FROM QUALITATIVE CRITICISM, TO EVALUATE AND SELECT REPRESENTATIONS WITH DOMAIN-INDEPENDENT METRICS, TO GENERATE AND SELECT PLAUSIBLE FEATURES, AND TO DISCOVER FEATURES VIA INTERESTINGNESS HEURISTICS DURING PROBLEM SOLVING.