Bayesian Flow Networks: A New Class of Generative Models

Explore a comprehensive lecture on Bayesian Flow Networks (BFNs), a novel class of generative models presented by Alex Graves at Valence Labs. Delve into the intricacies of this new approach, which modifies independent distribution parameters using Bayesian inference and neural networks. Learn how BFNs simplify the generative process compared to diffusion models, eliminating the need for a forward process. Discover the derivation of discrete and continuous-time loss functions for various data types, and understand how BFNs optimize data compression without architectural restrictions. Examine the model's competitive performance in image modeling and its superiority in character-level language modeling. Follow along as the speaker covers distributions, mathematical foundations, continuous and discretized data handling, pseudocode implementation, and concludes with an insightful Q&A session.

- Intro
- Discussion of distributions
- The Math
- Continuous Data
- Pseudocode
- Discretized Data
- Discrete Pseudocode
- Q&A