Pantelis Vafidis

Excited to share our work on disentangled/abstract representations, to appear at #ICLR2025 (@iclr_conf)! We mathematically prove and experimentally demonstrate that multi-task learning leads to disentangled representations, and propose a unifying mechanism for generalization in brains and machines: parallel processing (🧵+paper below) Our work connects to the Platonic representation hypothesis, suggests why alignment across models/organisms can occur, and shows why transformers excel at constructing world models 🤖🚀

Two of the most cracked ppl I’ve ever known building something that can truly make LLMs personalized. Congrats @ABhargava2000 and @witkowski_cam, excited to see where this goes!

@YinChaoqun @KordingLab Agreed. But this blogpost shows that the criterion used to detect “line attractors” is extremely lax: even random dynamical systems pass it 50% of the time, and get classified as “approximate line attractors”, which is clearly wrong.

@vafidisp @KordingLab I agree that the model somehow appears to fit everything. But I think it results from the very low-dimensional neural data itself. Its simple dynamics (mostly slow ramping) limits the model's complexity.

Attractors are usually not mechanisms.

@YinChaoqun @KordingLab Very problematic set of studies when it comes to quantification and modelling: aman-bhargava.com/dynamical/syst…

@KordingLab Isn't there a recent paper showing the existence of attractors in the mouse brain with causal experiments? nature.com/articles/s4158…

@rudzinskimaciej Yes, I could certainly see inhibition helping with decorrelation. Thanks for pointing it out. Not sure about the frequencies part though. Any references?

Columns seem to have few mechanisms that keep their representations different as a design choice both through inhibitory neurons and frequencies organization I'm pointing it as it's testable architecture choice that maps exactly to what you suggested Random initial weights wouldn't give different enough results, we know already that NN with random initiation but trained on the same data usually form similar representation

Yes great point, I forgot to reference the 1000 brains theory here but we do in the paper. One main difference is that they require dense signals to map the world, while we show that it can be done with sparse signals With regards to differentiation of cortical columns, it may simply come about due to the random initial projections (similar to heads in transformers or filters in CNNs)

If you're at #ICLR2025, and interested in how we can guarantee true out-of-distribution generalization in neural networks (extrapolation), Aman Bhargava (@ABhargava2000) and I will be presenting our work tomorrow Saturday the 26th at 3:00-5:30pm, at Hall 3 (poster number #69) We will be happy to see you there! short presentation + slides: iclr.cc/virtual/2025/p…

Finally, huge thanks to amazing collaborator Aman Bhargava (@ABhargava2000) for recognizing the mathematical potential of this project and doing the theory part, and advisor Antonio Rangel! This project a prime example of the amplifying effect of great collaborations. Looking forward to more! Link to top:

Thanks for reading this far! For an in depth view of the above, I include the paper below (it’s 40 pages long!). Tldr: it worked no matter what we threw at it! And if you happen to be in Singapore for #ICLR2025, we will be presenting at poster session 6 on Saturday the 26th, 3:30-5 pm (Hall 3 + Hall 2B #69). We will be happy to see you there! arxiv.org/abs/2407.11249

Excited to share our work on disentangled/abstract representations, to appear at #ICLR2025 (@iclr_conf)! We mathematically prove and experimentally demonstrate that multi-task learning leads to disentangled representations, and propose a unifying mechanism for generalization in brains and machines: parallel processing (🧵+paper below) Our work connects to the Platonic representation hypothesis, suggests why alignment across models/organisms can occur, and shows why transformers excel at constructing world models 🤖🚀