PhD Defense: Enriching Communication between Humans and AI Agents

Khanh Nguyen
06.08.2022 10:00 to 12:00

IRB 5137

Equipping AI agents with effective, human-compatible communication capabilities enables them to reliably serve and aid humans. On one hand, AI agents must be able to accurately infer intentions and extract knowledge from human utterances. On the other hand, they should also help humans comprehend and regulate their behaviors, by conveying their (un)certainties and proactively consulting humans when facing difficult situations.My talk presents new training and evaluation frameworks that enrich communication between humans and AI agents. I focus on improving two important capabilities of an agent: (1) the ability to learn from humans through language-based communication and (2) the ability to request and interpret information from humans to make better decisions. To improve the first capability, I study the possibility and challenges of training sequential decision-making agents with noisy simulated human ratings. Enabling humans to transfer knowledge to agents via richer communication media than numerical ratings, I propose a learning framework that allows agents to learn from descriptive human language. To enhance the second capability, I introduce new benchmarks that simulate and measure an agent’s ability to actively ask for and interpret information from humans to successfully complete tasks. On these benchmarks, I build agents that are capable of requesting rich, contextually useful information and show that they significantly outperform agents without such capability. I conclude with a discussion on future directions to endow AI agents with more expressive communication capabilities and models of cognition.
Examining Committee:

Chair:Dean's Representative:Members:

Dr. Hal Daumé III Dr. Philip Resnik Dr. Yonatan Bisk (CMU) Dr. Abhinav Shrivastava Dr. Pratap Tokekar