Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport

Explore the advancements in flow-based generative models through this comprehensive conference talk. Dive into the world of continuous normalizing flows (CNFs) and discover the innovative generalized conditional flow matching (CFM) technique. Learn how CFM overcomes limitations in simulation-based maximum likelihood training, offering a stable regression objective similar to diffusion models while maintaining efficient inference. Understand the advantages of optimal transport CFM (OT-CFM) in creating simpler, more stable flows for faster inference. Examine the application of these techniques in various conditional and unconditional generation tasks, including single cell dynamics inference, unsupervised image translation, and Schrödinger bridge inference. Gain insights into the comparison between diffusion models and CNFs, and explore the potential of training CNFs like diffusion models. Follow along as the speaker covers topics such as flow matching, conditional flow matching, the impact of probability path choice on flow properties, and the relationship between score and flow matching. Conclude with key takeaways and participate in a Q&A session to deepen your understanding of these cutting-edge concepts in AI and generative modeling.

- Intro
- Background on diffusion + flow models
- Why do diffusion models beat CNFs?
- Main idea: how can we train a CNF like a diffusion model?
- Flow matching
- Conditional flow matching
- Properties of flow depend on the choice of the probability path
- Score and flow matching
- Main takeaways
- Q+A