| Title | : | Learning Interpretable Models for Autonomous Assessment of Taskable AI Systems |
| Speaker | : | Pulkit Verma (MIT) |
| Details | : | Tue, 21 Jan, 2025 12:00 PM @ Online |
| Abstract: | : | Abstract:
Taskable AI systems are increasingly expected to operate in diverse, user-specific environments and perform tasks tailored to individual needs, all while improving over time. A critical challenge is determining whether these AI systems can safely and effectively fulfill the specific tasks users have in mind.
In this talk, I will present my work on addressing this challenge, including methods for conducting independent, post-deployment differential assessments of AI systems. I will discuss the necessary requirements these systems must meet to enable such assessments. A key aspect of my approach involves a personalized AI assessment module that allows the AI to execute instruction sequences in simulators and respond to user queries about these executions. Our findings demonstrate that even a basic query-response interface can efficiently produce a user-interpretable model of an AI system's capabilities. Speaker Bio: Pulkit Verma is a Postdoctoral Associate at the Interactive Robotics Group at the Massachusetts Institute of Technology, where he works with Prof. Julie Shah. His research focuses on the safe and reliable behavior of taskable AI agents. He investigates the minimal set of requirements in an AI system that would enable a user to assess and understand the limits of its safe operability. He received his Ph.D. in Computer Science from Arizona State University, where he worked with Prof. Siddharth Srivastava. Before that, he completed his M.Tech. in CSE at IIT Guwahati with Prof. Pradip K. Das. He was awarded the Graduate College Completion Fellowship at ASU in 2023 and received the Best Demo Award at the AAMAS 2022 Conference. His work has been published at top conferences including AAAI, NeurIPS, KR, and ICAPS. |
