Title: Interactive and Explainable Machine Learning Methods With Humans
Date: June 12
Time: 1:00 PM (Eastern)
Location:
- In Person at Klaus 2456 (Classroom Wing) or
- Virtual at https://gatech.zoom.us/j/99019110918?pwd=Vi9EMUM4bDBZWnNCczRQamNNcmtUQT09
Andrew Silva
Computer Science Ph.D. Candidate
School of Interactive Computing
Georgia Institute of Technology
Committee:
Dr. Matthew Gombolay (Advisor) – School of Interactive Computing, Georgia Institute of Technology
Dr. Sonia Chernova – School of Interactive Computing, Georgia Institute of Technology
Dr. Mark Riedl – School of Interactive Computing, Georgia Institute of Technology
Dr. Diyi Yang – Computer Science Department, Stanford University
Dr. Barry Theobald – Machine Learning Research, Apple, Inc.
Abstract:
This dissertation introduces and evaluates new mechanisms for interactivity and explainability within machine learning, specifically targeting human-in-the-loop learning systems. The contributions of this dissertation aim to substantiate the thesis statement: Interactive and explainable machine learning yields improved experiences for human users of intelligent systems.
The dissertation work will show that machine learning with human expertise offers improved performance in task success rates and reward, introducing a novel neural network architecture and an approach to goal-specification using language commands. I will then discuss how machine learning with explainability improves human perceptions of intelligent agents and enhances user compliance with agent suggestions, detailing technical contributions and a large-scale user study on perceptions of explainability mechanisms. Finally, I will overview my work in personalization for machine learning and the ways in which personalized machine learning enables improved performance for a large heterogeneous population of users. I offer both novel technical methods for interactivity and explainability within machine learning, as well as user studies to empirically validate my technical contributions. My dissertation will conclude with a presentation on recent work for personalizing explainability mechanisms to users in the task-oriented setting of guiding a simulated self-driving car in an unseen environment, navigating a tradeoff between participant-preference and task-performance.