Seminar • Artificial Intelligence • Learning Generative Models from a Control Perspective for Scientific Discovery

Wednesday, February 7, 2024 10:30 am - 11:30 am EST (GMT -05:00)

Please note: This seminar will take place in DC 1304.

Dinghuai Zhang, PhD candidate
Mila

Advancements in scientific discovery have always been at the forefront of human endeavor, particularly in complex domains such as molecule synthesis. The intrinsic challenges in these fields stem from two main factors: the vast and combinatorially complex high-dimensional search spaces, and the costly evaluation of scientific hypotheses. Therefore, leveraging machine learning offers a promising avenue to expedite the scientific discovery process.

In this talk, I will demonstrate how we can learn generative models to generate scientific proposals to efficiently traverse the candidate space. 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 control perspective. The algorithms I designed, including GFlowNet-based generative models, significantly outperform previous methods in discovering diverse and high-quality samples across multi-disciplinary applications such as protein design and combinatorial optimization. By enhancing exploration with structured probabilistic models, I believe my works will open new frontiers in large scale scientific discovery problems.


Bio: Dinghuai Zhang is a PhD candidate at Mila, advised by Prof. Yoshua Bengio and Prof. Aaron Courville. His research focuses on the intersection of generative modeling and scientific discovery. From a methodology perspective, he studies how to incorporate structured exploration into inference problems such as sampling, leveraging the power of the generative flow network (GFlowNet) framework which revolves around active learning, Bayesian inference, black box optimization, and reinforcement learning. He develops probabilistic methods for applications on different sorts of scientific discovery tasks spanning from biology, chemistry, and material science.

His contributions have been recognized through spotlight and oral presentations at top tier machine learning conferences such as NeurIPS, ICML, and ICLR. Dinghuai has got the Borealis Global AI fellowship, and has spent time in FAIR lab (Meta AI).  His work has been applied in world-class pharmaceutical companies such as Recursion for effectively searching new drugs in high dimensional candidate spaces. Dinghuai obtained a bachelor’s degree in math from Peking University.