RT @FrankNoeBerlin: Super excited to preprint our work on developing a Biomolecular Emulator (BioEmu): Scalable emulation of protein equili….

RT @su_lin_liu: Discrete generative models use denoisers for generation, but they can slip up. What if generation *isn’t only* about denois….

RT @json_yim: Combining discrete and continuous data is an important capability for generative models. To address this for protein design,….

@json_yim @valence_ai Big thanks to amazing co-lead @json_yim and our advisors @BarzilayRegina , @tom_rainforth , Tommi Jaakkola. 7/7.

@json_yim We are giving a talk about the work this Tuesday 11am EST/4pm GMT @valence_ai . Code for the pure discrete model: Code for protein co-design experiments: 6/7.

We combine our discrete flow with SE3 flow matching to make a protein structure-sequence generative model that achieves state-of-the-art protein generation results. See @json_yim 's post for more details. 5/7.

Our discrete flow framework is built using natural analogies to continuous space flow matching. The generative process is constructed by considering a mixture of conditional flows. We then sample using discrete rate matrices analogous to continuous space vector fields. 4/7

Training is with the standard cross-entropy loss -> no more complex continuous time discrete diffusion objectives!.We can tune sampling dynamics at inference time without any re-training, enabling superior performance to standard discrete diffusion. 3/7.

How do you define a flow on discrete data?. - We define the flow ODE in probability space, as a mixture of linear interpolations between noise and data. - We then find a continuous time markov chain that marginally matches this probability flow and sample it to generate data. 2/7.

New paper: how to do flow matching on discrete data. Flows give a simple generative framework and better performance than discrete diffusion models. Discrete flows are easily combined with continuous flow matching for multimodal models. . A thread 1/7

How can we apply diffusion models to data with varying dimensionality? We use jump diffusions to simultaneously generate the size and state values for varying size data e.g. molecules. w/ @willarvey @wh1lo @ValentinDeBort1 @tom_rainforth @ArnaudDoucet1

@JoeJBenton @ValentinDeBort1 @tom_rainforth @GeorgeDeligian9 @ArnaudDoucet1 Code and models available at

📰A Continuous Time Framework for Discrete Denoising Models. Operating in continuous time gives us higher performance generative samplers and error bounds on discrete spaces. w/ @JoeJBenton @ValentinDeBort1 @tom_rainforth @GeorgeDeligian9 @ArnaudDoucet1

I have written a blog post describing our use of Reinforcement Learning to create an online objective for sequential Variational Inference.🌐 📰

RT @ArnaudDoucet1: Online Variational Filtering and Parameter Learning with @AndrewC_ML @YuyangShi0 & @tom_rainfort….