iiinova — where your digital self keeps evolving

Jonas Köhler @jonkhler.argmin.xyz · 25/12/2024

Fun challenge that is apparently hard for LLMs (at least all those I could get my hands on): solving carefully drafted anagrams! New ARC challenge? :D PS: credits go to family Lehmann from Kassel who invent those ridiculously hard Xmas puzzles every year!

Jonas Köhler @jonkhler.argmin.xyz · 20/12/2024

Actually you convinced me to start one! <3

Jonas Köhler @jonkhler.argmin.xyz · 20/12/2024

I should... But all the time. Man there are so many things I would love to write a blog about and I never find the time 🫣

110

Jonas Köhler @jonkhler.argmin.xyz · 20/12/2024

Segment radixsort, prefixsums and pointwise transforms are just OPed on GPUs. The amount of throughput you can get on a single cheap device is ridiculous! And you can implement so much crazy stuff with just that. ❤️

190

Reposted by Jonas Köhler

Bálint Máté @balintmate.bsky.social · 17/12/2024

hello bluesky! we have a new preprint on solvation free energies: tl;dr: We define an interpolating density by its sampling process, and learn the corresponding equilibrium potential with score matching. arxiv.org/abs/2410.15815 with @francois.fleuret.org and @tbereau.bsky.social (1/n)

13310

Jonas Köhler @jonkhler.argmin.xyz · 13/12/2024

Well, in terms of adding to climate change both are not wrong 🫣

0110

Jonas Köhler @jonkhler.argmin.xyz · 06/12/2024

Congrats Frank + collaborators! :)

121

Jonas Köhler @jonkhler.argmin.xyz · 06/12/2024

Awesome work! Amazing to see this out :)

0100

indeed - including less stress/boredom for the audience :D

David Ho @davidho.bsky.social · 25/02/2024

Reposted by Jonas Köhler

“Sorry”

Underwater photograph of a large whale with a diver in the background, next to an article title that reads "WHAT'S THE FIRST THING WE SHOULD SAY TO THE WHALES?" by Ross Andersen.

522079390

Reposted by Jonas Köhler

Max Schambach @maxschambach.bsky.social · 02/12/2024

I’ll be at NeurIPS next week –happy to connect! Hit me up if you’d like to meet up and chat :) (tabular deep learning; in-context learning; SSL; AI in EU; politics; where to get good coffee/food; your favourite Parks and Rec moment; …)

051

Also regarding this whole insecurity business: one can simply say "good question, honestly I don't know, maybe someone in the audience can help us to clear this up?". I would prefer to hear that a lot more - especially from senior academics - rather than assertive statements.

I am personally not at all impressed, intimidated or even excited by a polished business school or flawless but boring diplomat's talk. I personally prefer having one or two memes and puns on the slides so that I don't sleep away rather than boring but correct formal stuff...

110

More on the serious note: my personal advice - especially for many academics - 1. be yourself (=don't act like sb) 2. don't take things too seriously (yes even job talks) 3. don't learn your slides by heart - use them as a visual guide to tell a friend what excites you.

230

Jonas Köhler @jonkhler.argmin.xyz · 01/12/2024

This reads like pickup artists giving presentation advice....

2120

Every two days I get this "invalid handle" label for my custom handle @jonkhler.argmin.xyz (using the DNS TXT config). This seems to be a bug - whenever I reset it, it's fine for another bunch of days. Anyone else experiencing this? If so, what's the cause/fix? #bluesky @jay.bsky.team

I guess reformulating the OT problem in a way that you minimize the total action along paths while matching the marginals would give you that.

010

And in the same way OT flow matching showed that there is more in these models than just getting a good marginal (here resulting in faster integration) you could also think about imposing physically amenable states all along the path.

100

That's all a bit hand wavy, but the idea would be: have equilibrium samples in A and B, learn transport maps between them, then study the paths learned to see where and how transitions actually happen.

100

Ah we chatted about applications for predicting trajectories instead of aiming for accurate T=0 images. In this context I remembered that folks looked into DMs to learn physically possible transition paths between two meta stable states using flow matching {...} + u(x) based constraints on paths.

Gabriel Peyré @gabrielpeyre.bsky.social · 30/11/2024

Reposted by Jonas Köhler

I wrote a summary of the main ingredients of the neat proof by Hugo Lavenant that diffusion models do not generally define optimal transport. github.com/mathematical...

523845

I read gossip that LoRa can bring down costs a lot! So maybe not that much? en.m.wikipedia.org/wiki/LoRa

en.m.wikipedia.org

LoRa - Wikipedia

Sander Dieleman @sedielem.bsky.social · 30/11/2024

Reposted by Jonas Köhler

The link between diffusion models and optimal transport is still a bit of an enigma to me. One thing that's clear: different diffusion models trained on similar datasets tend to recover similar mappings. If these are generally not OT, in what sense are they optimal instead?

311411

Oh yeah, that's on my reading list for a while 🫣🫣

010

More generally: lots of work in ML focuses on the result of the process (final images) and invariance under initialisation seeds. All interesting enough :) But the value of the actual trajectories is underexplored imho. There surely is tons of work already but still seems a bit niche.

Dinner table convo from yesterday: if we go from standard diffusion models to flow-matching/Schrödinger bridges and simply try to map two (general) densities onto. What kind of (stochastic) paths do we learn between them? And if it's all physical are those paths too? And if not how to enforce it?

230