Academic

Sampling in high-dimensional spaces presents a considerable challenge in various areas of statistical computation. Traditional methods like Markov Chain Monte Carlo (MCMC) often struggle with convergence and computational inefficiency in complex settings. Variational inference provides an alternative through amortization but also suffers from limitation of variational families and issues in optimization.  In this talk, I will demonstrate how we can approach probabilistic inference problems from a control perspective. The key insight of my research is framing the data generation process of structured scientific objects as a sequential decision-making process, and learning the generation policy from a reinforcement learning perspective. The resulting framework, generative flow network (GFlowNet), provides a novel way of conducting amortized inference where systematic improvements in exploration are achievable. The algorithms I designed significantly outperform previous methods in capturing diverse modes in high-dimensional spaces. My work not only advances the methodology in statistical computation, but also has substantial applied implications for meaningful scientific discovery in areas such as molecule synthesis and protein design.