Sam Gijsen

ML @ Tübingen AI Center, Hertie AI

[Sam Gijsen @SamCJG]

Can diffusion transformers do in-silico neuroscience?

In a new preprint, we train models to generate neural (fMRI) time series. We condition per time step by injecting conditioning tokens directly in context. We evaluate on hundreds of unseen task conditions.

Some results! 🧵

[Sam Gijsen @SamCJG]

@giffmana They know.

[Sam Gijsen @SamCJG]

@DamienTeney @giffmana did you go into the maze?

[Sam Gijsen @SamCJG]

@wyqtor @francoisfleuret RSI banned, everyone go home

[Sam Gijsen @SamCJG]

This constitutes just a first step towards in silico experimental design but already works surprisingly well, with tons of open paths for improvements!

Paper: https://t.co/3727sjQc1u

[Sam Gijsen @SamCJG]

Can diffusion transformers do in-silico neuroscience?

In a new preprint, we train models to generate neural (fMRI) time series. We condition per time step by injecting conditioning tokens directly in context. We evaluate on hundreds of unseen task conditions.

Some results! 🧵

[Sam Gijsen @SamCJG]

We quantify training support distance in terms of both spatial activation maps and language embeddings of task descriptions, both showing distinct degradation profiles.

[Sam Gijsen @SamCJG]

@HumanProgress 150K deaths per 100K people? thats tough

[Sam Gijsen @SamCJG]

@iamgingertrash @Tom_A_Lynch @boneGPT That reads to me like a well functioning hedging market, at which point you're paying to decrease vol which is akin to using an inflating currency? Bit skeptical of choosing to stabilize against what is pretty much the least fungible / most volatile / annoying-to-store thing

[Sam Gijsen @SamCJG]

@Tom_A_Lynch @iamgingertrash @boneGPT Why? we already have evidence today that currency volatility limits MOE adoption (and inversely, the success of stables)

[Sam Gijsen @SamCJG]

@iamgingertrash @jefernandez8 SPX? ie a 3% move?

[Sam Gijsen @SamCJG]

@giffmana @ens_pyrz Gaben pls allow the TI prize to be won by AI?

[Sam Gijsen @SamCJG]

@iamtrask @jeremyphoward Why is running all of the biggest models instead of just 1 good for efficiency?

[Sam Gijsen @SamCJG]

@giffmana Can’t blame river for this one

[Sam Gijsen @SamCJG]

@MParakhin Find it pretty hard to get 5.5 pro to put in enough effort; things always come back undercooked

[Sam Gijsen @SamCJG]

@iScienceLuvr I much prefer codex responses in general, they tend to not be as verbose and don’t have this “to conclude … what I’d actually do … your best option … to summarise” awkward structure that the chat models have