Excited to share Flow Matching Policy Gradients: expressive RL policies trained from rewards using flow matching. It’s an easy, drop-in replacement for Gaussian PPO on control tasks.
Decentralized Diffusion Models power stronger models trained on more accessible infrastructure.
DDMs mitigate the networking bottleneck that locks training into expensive and power-hungry centralized clusters. They scale gracefully to billions of parameters and generate
I’ll be at #NeurIPS2024 this week presenting Rethinking Score Distillation as a Bridge Between Image Distributions!
Poster Presentation: Friday 4:30-7:30 PM
Come chat with me or @holynski_ about lifting diffusion models to 3D!
our new system trains humanoid robots using data from cell phone videos, enabling skills such as climbing stairs and sitting on chairs in a single policy
(w/ @redstone_hong@junyi42@davidrmcall)
Decentralized Diffusion Models power stronger models trained on more accessible infrastructure.
DDMs mitigate the networking bottleneck that locks training into expensive and power-hungry centralized clusters. They scale gracefully to billions of parameters and generate
RDM is now published at Nature Methods! This was a 3 year effort and my introduction to academic research. I’m fortunate to have been mentored by one of the smartest people I’ve ever met @the_legitamit!
Ring deconvolution microscopy is now published at @naturemethods!
nature.com/articles/s4159…
There are some fun new additions including light-sheet deconvolution 🫡
Stay tuned for the official python package release next week! Any feature suggestions are more than welcome 😃
Standard diffusion models communicate gradients at every optimization step, a network load that only centralized clusters can support. Decentralized diffusion models divide training into independent pieces that can proceed on different hardware in different locations. This is a
This was only possible thanks to mentors Matthew Tancik and Jiaming Song (@baaadas), as well as the support of the rest of the fantastic @LumaLabsAI research team! Thank you for hosting me as an intern.
This produces an ensemble whose predictions combine at test-time. You can drop experts at test-time to save on computation. In fact, we only inference a single expert per step in our comparisons and we outperform standard diffusion models for with the same FLOP budgets at train