Lightnews — Scholar-powered news

Reposted by Paul Hagemann

Simon Olsson @smnlssn.bsky.social · May 27

We are looking for someone to join the group as a postdoc to help us with scaling implicit transfer operators. If you are interested in this, please reach out to me through email. Include CV, with publications and brief motivational statement. RTs appreciated!

1 8 14

Paul Hagemann @yungbayesian.bsky.social · May 26

what is so misunderstood about (3)?

2

Reposted by Paul Hagemann

Simon Olsson @smnlssn.bsky.social · May 6

2025 CHAIR Structured Learning Workshop -- Apply to attend: ui.ungpd.com/Events/60bfc...

5 13

Paul Hagemann @yungbayesian.bsky.social · Apr 25

cool work

Paul Hagemann @yungbayesian.bsky.social · Mar 19

best of luck marvin :)

1 2

Paul Hagemann @yungbayesian.bsky.social · Feb 27

ist auch logisch, dass greenpeace/foodwatch bspw den grünen näherstehen, da die ja die themen bespielen. neutralität wäre da ja eher lächerlich

Paul Hagemann @yungbayesian.bsky.social · Feb 18

was ist an der studie falsch

Paul Hagemann @yungbayesian.bsky.social · Feb 18

interesting point, but i would say (true) memorization is mathematically impossible. the underlying question is what generalization means when we are given finite training samples. it depends on the model and how long you train, see proceedings.neurips.cc/paper_files/... and arxiv.org/abs/2412.20292

Score-Based Generative Models Detect Manifolds

proceedings.neurips.cc

Paul Hagemann @yungbayesian.bsky.social · Feb 7

yes i agree, but for diffusion such a constant velocity/score field does not even exist

1 1

Paul Hagemann @yungbayesian.bsky.social · Feb 6

so in diffusion models the time schedule is so that we cannot have straight paths velocity fields (i.e., v_t(x_t) is constant in time), as opposed to flow matching/rectified flows where it is possible to obtain such paths (although it requires either OT/rectifying...)

1

Paul Hagemann @yungbayesian.bsky.social · Jan 23

yes lol thank you!

Paul Hagemann @yungbayesian.bsky.social · Jan 23

Check out our github and give it a try yourself! Lots of potential in exploring stuff like this also to other domains (medical imaging, protein/bio stuff)!

github.com/annegnx/PnP-...

Also credit goes to my awesome collaborators Anne Gagneux, Sego Martin and Gabriele Steidl!

github.com

1

Paul Hagemann @yungbayesian.bsky.social · Jan 23

Compared to diffusion methods, we can handle arbitrary latent distributions and also get (theoretically) straighter paths! We evaluate on multiple image datasets against flow matching+diffusion+standard PnP based restoration methods!

1

Paul Hagemann @yungbayesian.bsky.social · Jan 23

Our algorithm proceeds as follows: we do a gradient step on the data fidelity, reproject onto the flow matching path and then denoise using our flow matching model. This is super cheap to do!

1

Paul Hagemann @yungbayesian.bsky.social · Jan 23

Therefore, we use the plug and play framework and rewrite our velocity field (which predicts a direction) to instead denoise the image x_t (i.e., predict the MMSE image x_1). Then we obtain a "time" conditional PnP version, where we solve do the forward backward PnP at the current time and reproject

1

Paul Hagemann @yungbayesian.bsky.social · Jan 23

Our paper "PnP-Flow: Plug-and-Play Image Restoration with Flow Matching" has been accepted to ICLR 2025. Here a short explainer: We want to restore images (i.e., solve inverse problems) using pretrained velocity fields from flow matching. However, using change of variables is super costly.

1 5 15

Paul Hagemann @yungbayesian.bsky.social · Jan 1

very nice paper, only had a quick glimpse, but another aspect is that the optimal score estimator explodes if we approach t -> 0, which NNs ofc cannot replicate. how does this influence the results?

1 1

Paul Hagemann @yungbayesian.bsky.social · Nov 28

you might be onto sth haha

1

Paul Hagemann @yungbayesian.bsky.social · Nov 28

i guess the adam paper is a pretty good indicator how much ml papers are being published. looks like we are saturating since 2021

1 2

Paul Hagemann @yungbayesian.bsky.social · Nov 24

same experience here. i am not sure we need actual conference reviewing at all. why do we not all publish on openreview and if i use your paper/build upon/read it, i can write my opinion on it? without the accept reject stamp.

Paul Hagemann @yungbayesian.bsky.social · Nov 20

Here, one can see FID results for different beta! Indeed it seems to be fruitful to restrict mass movement in Y for class conditional cifar! We apply this also to other interesting inverse problems, the article can be found at arxiv.org/abs/2403.18705

1

Paul Hagemann @yungbayesian.bsky.social · Nov 20

We want to approximate this distance with standard OT solvers, and therefore introduce a twisted cost function. With this at hand, we can now do OT flow matching for inverse problems! The factor beta controls how much mass leakage we allow in Y.

1 2

Paul Hagemann @yungbayesian.bsky.social · Nov 20

This object has already been of some interest, i.e., it pops up in the theory of gradient flows. It generalizes the KL property quite nicely, and unifies some ideas present in conditional generative modelling. For instance, its dual is the loss usually used in conditional wasserstein gans.

1

Paul Hagemann @yungbayesian.bsky.social · Nov 20

Now does the same hold for the Wasserstein distance? Unfortunately not, since moving mass in Y-direction can be more efficient for some measures. However, we can fix that if we restrict the suitable couplings to ones, that only move mass in Y-direction.

1

Paul Hagemann @yungbayesian.bsky.social · Nov 20

In a somewhat recent paper we introduced conditional Wasserstein Distances. They generalize a property that basically explains why KL works well for generative modelling, the chain rule of KL!
It says that if one wants to approximate the posterior, one can also minimize the KL between joints.

1 15