Derivative-Free Guidance in Continuous and Discrete Diffusion Models

Explore a comprehensive talk on derivative-free guidance in continuous and discrete diffusion models presented by Xiner Li and Masatoshi Uehara from Valence Labs. Delve into the challenges of optimizing downstream reward functions while maintaining the naturalness of design spaces in diffusion models for images, molecules, DNA, RNA, and protein sequences. Learn about a novel iterative sampling method that incorporates soft value functions into pre-trained diffusion models' inference procedures, eliminating the need for differentiable proxy models or expensive fine-tuning. Discover how this approach enables direct utilization of non-differentiable features and reward feedback, making it applicable to recent discrete diffusion models. Examine the effectiveness of this algorithm across various domains, including image generation, molecule generation, and DNA/RNA sequence generation. Gain insights into the latest advancements in AI for drug discovery and connect with the speakers through the Valence Labs Portal community.

Derivative-Free Guidance in Continuous and Discrete Diffusion Models | Xiner Li and Masatoshi Uehara