Qiwei Di

Welcome to my homepage! My name is Qiwei Di. I’m a third-year CS PhD student at UCLA, fortunate to be advised by Prof. Quanquan Gu. Previously, I graduated from the math and applied math department in Tsinghua University.

Research Interests

My research interests lie in the field of machine learning, especially learning theory and theory-guided applications. I’m working on
Reinforcement Learning: Markov Decision Process, Dueling Bandits, Complexity Measures
Diffusion Models: Faster ODE/SDE Solvers, Unified Convergence Analysis, Theory of Optimal Transport
LLM: RLHF, LLM Robustness, Retrieval Augmented Generation (RAG)

Recent News

A new paper on unified convergence analysis of ODE solvers:
Unified Convergence Analysis for Score-Based Diffusion Models with Deterministic Samplers Runjia Li, Qiwei Di, Quanquan Gu, Preprint.

• We develop a key technical tool which enables us to bound the time derivative of the total variation (TV) distance between the final states of two ODE processes, through the difference in their drift terms and the divergence. As a direct application, we establish convergence guarantees for the continuous-time reverse ODE in the case of the OU forward process.
• We provide a unified convergence analysis framework for diffusion models with deterministic samplers.
• To demonstrate the generality and effectiveness of our framework, we apply it to two typical diffusion model settings, Variance Preserving (VP) forward process with exponential integrator (EI) numerical scheme, and Variance Exploding (VE) forward process with the DDIM numerical scheme.