CAIM talk by Alex Tong on machine learning and computational neuroscience
Date: June 24, 2026, 12:30 pm
Location: Medical University of Vienna, Anna Spiegel Research Building, Seminar Room Level 3
Speaker: Alex Tong
Title: Fast, Accurate, and Discrete: Next-Generation Flow Matching and Distillation
Abstract: Generative models have achieved remarkable success, but their reliance on iterative, multi-step numerical integration remains a fundamental bottleneck for inference. This talk outlines a trajectory of recent work aimed at drastically accelerating these models—reducing sampling steps and expanding their reach across both continuous and discrete domains.
Starting with Conditional Flow Matching (CFM) and minibatch optimal transport, we establish a simulation-free framework for training continuous normalizing flows. Building on this foundation, we explore techniques to push past traditional integration limits, including fast likelihood estimation (FALCON), rapid sampling of SDE paths (SSFM), and few-step generation via Distribution Matching Distillation (DMD). Finally, we move beyond continuous variables to introduce Coupling Models, demonstrating how direct couplings enable state-of-the-art, single-step generation for discrete domains like language modeling and biological sequence design.
Bio: Alexander Tong is a Principal Investigator at the Aithyra Research Institute for Biomedical Artificial Intelligence in Vienna, where he leads a research group focused around generative modeling for the life sciences. He earned his PhD in computer science from Yale University in 2021 under the advisement of Smita Krishnaswamy, followed by postdoctoral research at Mila mentored by Yoshua Bengio. Alongside his academic research, he co-founded Dreamfold, a company focused on generative models for protein design. He holds a BS and an MS in computer science from Tufts University.
This is part of the CAIM Talks series.