I visited La Brea Tar Pits this weekend, and captured this photo of what I think is the epitome of human curiosity: when encountering a mysterious black substance oozing out of the ground, we just can't help but poke it with a stick.

Conditional generation of graphs and molecules is a critical problem, but it's hard to do that with diffusion models. Gabriele Scalia and I developed a framework for graph diffusion where generation is fully interpretable and controllable. arxiv.org/abs/2302.03790

Last week, Alex Tseng presented his work on branched diffusion models. Very interesting method, I personally found very cool the idea of generation by analogy, called "Transmutation". Thanks Alex for the fantastic presentation 👏! You can watch it here youtube.com/watch?v=56111r…

Theory vs practice: diffusion-model performance depends on the diffusion kernel's prior. In our short paper from Genentech, we explore this dependency. Surprisingly, the optimal prior is not obvious, unlike what previous studies may have hinted at! arxiv.org/abs/2306.02957

Thank you to the WCB organizers and all the amazing work presented here! To be voted one of the most popular posters was truly an honor!

It was wonderful to introduce our work on mechanistically interpretable neural networks for regulatory DNA at MLGenX Workshop @ ICLR 2024. The workshop was incredible and we got so much excellent positive and constructive feedback! openreview.net/forum?id=O3cxF… #ICLR2024