Teemu Sarapisto

CS/ML PhD research in high-dim time-series @helsinkiuni

Before: 5y of C++/JS/VR/AR/ML at Varjo, Yle, Reaktor, Automattic

After dark: synthesizers & 3D gfx

[Teemu Sarapisto @Tsarpf]

New paper: Subsystem Discovery in High-Dimensional Time-Series Using Masked Autoencoders

Presented at the European Conference on Artificial Intelligence (ECAI) 2024

Code/data/paper: https://t.co/mJfFOWgJKK

Map graph learned from weather time-series, adjacency from 7 engines!

[Teemu Sarapisto @Tsarpf]

@dwarkesh_sp A good example of why I always love the guests, but not the hype from the host.

3b1b is trying to make the point that math is unique in how self-contained the system is, so automating it could actually work.

Then you turn the convo into the most obvious hyperbole :(

[Teemu Sarapisto @Tsarpf]

@MuteDialog @TrackingActions Takens' theorem only considers reconstructing the shape, up to an arbitrary diffeomorphism, of e.g. a 5D system from 1D observations. It does not directly discuss noisy observations, affine identifiability, or symbolic recovery of the governing equations.

[Teemu Sarapisto @Tsarpf]

@TrackingActions I've also been thinking about "multi-view identifiability" but for time-series/dynamics, so this is doubly interesting.

My angle has been "same device but in different environments" and what could you do with that prior knowledge, and could it help with identifiability?

[Teemu Sarapisto @Tsarpf]

@TrackingActions Hey, very cool continuation! I'm a big fan of CEBRA+DynCL. My Top-3 favorite works.

My #1 worry: this & DynCL theory use noise shape to fix the gauge. Say a pendulum: Euclidean vs angle/speed coordinates, what distinguishes them? Gaussian noise in only one gauge feels unlikely?

[Teemu Sarapisto @Tsarpf]

@MikaelNiku Munkin työpäivää piristää kun ikkunan vieressä punarintavanhemmat roudaa laudanrakoon matoa sisään ja kakkaa ulos ihan jatkuvalla syötöllä. Oiskohan kolme poikasta näkynyt ja kova härdelli. Kätevän näköistä kun näköjään poikaset osaavat itse ojentaa valmiin kakkapaketin ovelle😄

[Teemu Sarapisto @Tsarpf]

@klindt_david https://t.co/TAaWXtBF2C

They set up a system for searching for interesting sec bugs with agents. Given the same code as input, many LLMs seem to find the bugs (almost) as well.

So it's mainly not the model that changed, this is just an excellent ad campaign for a new model?

[Teemu Sarapisto @Tsarpf]

@ryanpirl https://t.co/7CweSBMHC4

JEPA/contrastive/siamese architectures have some theoretical backing that they recover the shape of the "true" latent space. Here used for mice.

[Teemu Sarapisto @Tsarpf]

This was a nice, non-technical, text trying to answer "what is mathematics?" in a philosophy of science kind of way. Recommended!

[Steven Strogatz @stevenstrogatz]

With the rise of AI in mathematics, many of us are rethinking what math is for—and what it even is. These are philosophical questions, and we should discuss them carefully and openheartedly. This classic essay by Reuben Hersh is a great place to start. https://t.co/Hq2NAe5mFh

[Teemu Sarapisto @Tsarpf]

@jerrrrrrryyyyy @typedfemale If real-world means in "production", then I agree 😁

[Teemu Sarapisto @Tsarpf]

@typedfemale Things like swiss roll/interlocked rings.

Also number of (near) zero eigenvalues in graph Laplacian = number of disconnected groups in graph = useful spectral graph stuff.

https://t.co/VxCB9Kwkwx this is what I've read and found useful, I'm not a mathematician 😁

[Teemu Sarapisto @Tsarpf]

@martinmbauer As opposed to going to a bad university? I wonder who makes/runs those.

[Teemu Sarapisto @Tsarpf]

@RichardDevine https://t.co/zHzsZKeMb0

Here's a youtube thing of it...

[Teemu Sarapisto @Tsarpf]

@RichardDevine Reminds me of when I did a mesh visualization like this, except for live audio playing on Windows (just whatever was playing in my Foobar/winamp/spotify) way back. Time flies, apparently it was 10 years ago 😄

https://t.co/FcssckV3e0

[Teemu Sarapisto @Tsarpf]

An LLM trained with text only until year 1931 is evaluated in its ability to learn to code at inference time 😄 Really nice experiment.

[David Duvenaud @DavidDuvenaud]

@geoffreyirving We tried that! The vintage models can just barely start to do simple things with Python, purely from in-context learning:

[Teemu Sarapisto @Tsarpf]

@jeremyphoward The 20usd claude subscription is totally useless for coding anyway: it runs out of quota before finishing a single thing, sometimes even before getting started at all, when doing longer planning.

So actually I say it is good if they stop selling it as if it was usable.

[Teemu Sarapisto @Tsarpf]

@fchollet lol nice trolling.

I used PyTorch for 4y, now 3y of Jax, and coming from functional programming background, I love it!

But for 1st time I'm finding something missing in Jax now: I want to run diffeq solvers in parallel *on the same GPU", and no streams in Jax. Pytorch has them!

[Teemu Sarapisto @Tsarpf]

@cs_serdar @fchollet Pretty crazy. I tried HF and lightning-whatever once and felt so not-in-control I got frustrated in minutes. Esp. now with agents doing the boilerplate I see 0 reason to use libs that just hide the things you need to change for ML research anyway . Jax+Equinox is 🔥

[Teemu Sarapisto @Tsarpf]

@WKCosmo They also got stuck in some not-so-chaotic loops quite often.

[Teemu Sarapisto @Tsarpf]

@WKCosmo The only problem the vibe-coded pendulums kept having is that even with supposedly symplectic solvers, they kept bleeding energy over very long trajectories.