Sam McCallum

PhD student in Mathematics at the University of Bath.

[Sam McCallum @sgmccallum]

@ZhengyangGeng We seem to be interested in similar things :)

You can get exact DEQ gradient in O(1) memory with a reversible fixed point solver.

https://t.co/EM7RYUMM7i

[Sam McCallum @sgmccallum]

@FrancoisRozet @Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath Yep that's correct for Algorithm 1 since the t > s diffusion contribution is fixed by the Euler-Maruyama step.

In general, for the instantaneous (s=t) matching + distillation objective, you need a g-network to inject the noise.

[Sam McCallum @sgmccallum]

@FrancoisRozet @Niklas_TR @zwblasingame @AlexanderTong7 @JamesFosterBath You're right that the f-network has capacity to learn the Ito map. But you need the g-network to be non-zero for t > s so that noise is injected into the step.

[Niklas Rindtorff @Niklas_TR]

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://t.co/Hy8WWZOnjE
Code: https://t.co/PMe6RoqyZA