The open-source method generates 3D molecular conformers.
Michael Bronstein#202
Introducing Strong Stochastic Flow Maps
TLDR: Stochastic Flow Maps where we learn the stochastic solution path.
Work led by Sam McCallum, @zwblasingame, with Timothy Herschelll, @AlexanderTong7, and @JamesFosterBath
Arxiv: https://arxiv.org/pdf/2606.01086 Code: https://github.com/sammccallum/ssfm
Users praised Strong Stochastic Flow Maps for stochastic paths as awesome and promising work because it explicitly models trajectories and raises exciting new questions.
No Digg Deeper questions have been answered for this story yet.
This looks neat! Flow maps are deterministic by default. You can make them stochastic by conditioning on a noise variable (e.g. meta flow maps, diamond maps). But if you want to get the exact same samples with fewer steps, you need to condition on the entire noise trajectory!
Introducing Strong Stochastic Flow Maps
TLDR: Stochastic Flow Maps where we learn the stochastic solution path.
Work led by Sam McCallum, @zwblasingame, with Timothy Herschelll, @AlexanderTong7, and @JamesFosterBath
Arxiv: https://arxiv.org/pdf/2606.01086 Code: https://github.com/sammccallum/ssfm
Shoutout to @PatrickKidger's diffrax. Tagging people that might be interested @StefanoErmon @volokuleshov @adityagrover_ @ssahoo_ @sedielem @karsten_kreis @RickyTQChen @sitanch @gottapatchemall @ZhengyangGeng @ArashVahdat @emiel_hoogeboom @TimSalimans @ValentinDeBort1 @poolio @jaschasd @nmboffi @msalbergo @peholderrieth @k_neklyudov @mmbronstein @PPotaptchik @DidiKieran @LucaEyring @HannesStaerk @GabriCorso
@Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath bruh we wrote the same paper lol. nice
@msalbergo @zwblasingame @AlexanderTong7 @JamesFosterBath I will turn this tweet into a Slack sticker
Diffusion models are powerful generative models, but running them can be slow because we're taking many small stochastic steps from noise to data. Stochastic Flow Maps allow us to take fewer steps - which speeds up sampling.
However, current approaches learn only the marginal distributions of the process (weak convergence), rather than the solution path itself (strong convergence). This matters if we want to effectively steer our model during sampling:
This approach raises a couple new, exciting, questions: For example, can we apply this approach to learn Stochastic Flow Maps of Underdampened Langevin Dynamics - to speed up the current workhorse of molecular simulations? Can we use this approach to run protein binder design more effectively? Can we extend this approach to other consistency objectives, such as Lagrangian and Eulerian?
For example, we might want to do a few-step lookahead with respect to our current diffusion trajectory. Currently, this is done by using deterministic flow maps, which do not adhere to the very same diffusion trajectory we're on. In contrast, Strong Stochastic Flow Maps adhere to the path and recover approximately the same denoised state.
Strong Stochastic Flow Maps learn the solution path itself by being conditioned on polynomial approximations of the brownian noise along the path. These polynomials also allow us to simply construct a semigroup consistency objective: The coefficients over the time intervals [s, u] and [u, t] compose into the coefficients for the time interval [s, t].
@Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath haha, hard to make into a square icon
@Niklas_TR @PatrickKidger @StefanoErmon @volokuleshov @adityagrover_ @ssahoo_ @sedielem @karsten_kreis @RickyTQChen @sitanch @gottapatchemall @ZhengyangGeng @ArashVahdat @emiel_hoogeboom @TimSalimans @ValentinDeBort1 @poolio very cool!!
@Niklas_TR @PatrickKidger @StefanoErmon @adityagrover_ @ssahoo_ @sedielem @karsten_kreis @RickyTQChen @sitanch @gottapatchemall @ZhengyangGeng @ArashVahdat @emiel_hoogeboom @TimSalimans @ValentinDeBort1 @poolio Great work!
@Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath Interesting work. Explicitly modeling stochastic trajectories rather than focusing solely on endpoints feels like a promising direction for generative modeling.
@Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath Awesome work!!