Andrew Campbell

Very excited to share our preprint: Self-Speculative Masked Diffusions We speed up sampling of masked diffusion models by ~2x by using speculative sampling and a hybrid non-causal / causal transformer arxiv.org/abs/2510.03929 w/ @ValentinDeBort1 @thjashin @ArnaudDoucet1

Super excited to preprint our work on developing a Biomolecular Emulator (BioEmu): Scalable emulation of protein equilibrium ensembles with generative deep learning from @MSFTResearch AI for Science. #ML #AI #NeuralNetworks #Biology #AI4Science biorxiv.org/content/10.110…

Discrete generative models use denoisers for generation, but they can slip up. What if generation *isn’t only* about denoising?🤔 Introducing DDPD: Discrete Diffusion with Planned Denoising🤗🧵(1/11) w/ @junonam_ @AndrewC_ML @HannesStaerk @xuyilun2 Tommi Jaakkola @RGBLabMIT

@FrankNoeBerlin Thanks Frank! This is a choice we can make at inference time, we can sample with 'low stochasticity' where tokens commit or with 'high stochasticity' where tokens can switch back to mask (see gif). This enables the model to correct mistakes during generation.

@AndrewC_ML This looks really cool, congrats Andrew. In the movie it looks like once the residues commit to one amino acid they don’t change anymore. Is that expected?

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. arxiv.org/abs/2402.04997 A thread 1/7

Combining discrete and continuous data is an important capability for generative models. To address this for protein design, we introduce Multiflow, a generative model for structure and sequence generation. Preprint: arxiv.org/abs/2402.04997 Code: github.com/jasonkyuyim/mu… 1/8

@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 portal.valencelabs.com/events/post/im… Code for the pure discrete model: github.com/andrew-cr/disc… Code for protein co-design experiments: github.com/jasonkyuyim/mu… 6/7

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 arxiv.org/abs/2305.16261 w/ @willarvey @wh1lo @ValentinDeBort1 @tom_rainforth @ArnaudDoucet1

@JoeJBenton @ValentinDeBort1 @tom_rainforth @GeorgeDeligian9 @ArnaudDoucet1 Code and models available at github.com/andrew-cr/tauL…

📰A Continuous Time Framework for Discrete Denoising Models Operating in continuous time gives us higher performance generative samplers and error bounds on discrete spaces. arxiv.org/abs/2205.14987 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 🌐 andrew-cr.github.io/posts/RLandVI/ 📰 arxiv.org/abs/2110.13549

Online Variational Filtering and Parameter Learning arxiv.org/abs/2110.13549 with @AndrewC_ML @YuyangShi0 & @tom_rainforth accepted as Oral #NeurIPS2021. We exploit forward Bellman recursions for the ELBO & its gradients in state-space models and RL ideas.