A Generative Framework for Designing Interactions to Overcome the Gaps between Humans and Imperfect AIs Instead of Improving the Accuracy of the AIs

Abstract

My research focuses on improving human-machine collaboration in the context of machine learning, particularly by recognizing the limitations and potential for errors in machine learning techniques and designing effective interactions for filling the gaps between humans and them. To this end, I have explored the application of machine learning in a variety of domains, such as malware analysis, music recommendation, conversation analysis, photo editing, and video-based learning. I also worked on clarifying the limitations of the current technologies by using adversarial approaches and qualitative methods. My thesis is planned to synthesize what I learned from these projects into design principles for constructing interactions that take full advantage of imperfect machine learning models. I particularly put emphasis on deriving principles that do not depend on the fine-tuning of the models, thereby providing a generative framework allowing researchers and practitioners to design a range of effective intelligent interactions without incurring significant computational and data collection costs.