Title:  Information-theoretic Interactive Machine Learning

Committee: 

Dr. Rozell, Advisor       

Dr. Davenport, Chair

Dr. Bloch

Abstract:

The objective of the proposed research is to frame interactive machine learning (IML) as a coding problem and directly apply principles and algorithms from information theory to design and analyze high-performing IML systems that efficiently utilize interactions between the labeling expert and learner. As a step towards integrating concepts from information theory into IML, we directly apply recently developed optimal feedback-coding schemes to sequential human-computer interaction systems in tasks including brain-computer interfacing and interactive object segmentation. We then transition to develop novel information maximization methods for efficient query selection in interactive similarity embedding construction and preference learning systems, using computationally efficient information maximization strategies. Finally, we propose to utilize lessons from our first two aims to enhance a systems-level understanding of IML by analyzing the information content conveyed in interactive systems, leading to new algorithmic and theoretical insights into optimal IML system operation.