Towards Efficient AI Training: Theoretical Foundations of Sample and Computational Efficiency in Decision-Making

Speaker: Jian Qian, MIT

Title: Towards Efficient AI Training: Theoretical Foundations of Sample and Computational Efficiency in Decision-Making

Date: Thursday, 27 March 2025

Time: 10:00am - 11:00am

Join Zoom Meeting: https://hkust.zoom.us/j/96688516988?pwd=qfj1PQIjEi0I75lwVGfY7PurdPDRBW.1
Meeting ID: 966 8851 6988
Passcode: 202526

Abstract:

AI-powered decision-making systems have revolutionized many fields by transforming complex data into actionable insights. However, challenges remain in optimizing their training efficiency in terms of samples and computation required.

To address these challenges effectively, it is essential to develop solutions grounded in rigorous theoretical principles. In this talk, I will address two critical aspects of AI design that respond to these challenges: (i) Sample Efficiency: I will introduce a general framework for interactive decision-making, encompassing multi-armed bandits and reinforcement learning, which are core models behind systems like AlphaGo and large language models (LLMs). I will present the learning limits and a general algorithm design principle under this framework, which enables the development of new algorithms with provable near-optimal guarantees. (ii) Computational Efficiency: I will revisit the foundational assumption of smoothness, which enables provably efficient optimization methods in machine learning. After demonstrating the impracticality of this assumption in real-world scenarios, I will introduce a generalized concept of smoothness that takes a step toward bridging the gap between optimization theory and practical machine learning applications.

By providing these theoretical insights, this talk will deepen our understanding of AI's fundamental limits and help guide the development of systems that are both sample-efficient and computationally efficient.

Biography:

Jian Qian is a Ph.D. student in the Department of Electrical Engineering and Computer Science at MIT, advised by Prof. Alexander Rakhlin. His research focuses on understanding the foundamental aspects of machine learning and decision making. His work has been published in top-tier ML conferences. He is also enthusiastic about exploring new research directions that bridges the theory and practice of machine learning.