🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

ICML 2/2 🤩 Charlie O'Neill et al show with theory/experiments: sparse coding > SAEs arxiv.org/abs/2411.13117 Patrik Reizinger et al argue for empirical identifiability theory arxiv.org/abs/2504.13101 Congrats on great science & big thanks to AC+reviewers for their time and effort 🙏

1/🧵 Thrilled to announce our ICML‑accepted work (done in my undergrad days with @david_klindt and Alim Gumran) that questions the linear‑only dogma in sparse autoencoders👇

Science is a long arduous process of critical thinking, hypothesis testing and iteration that, maybe sometimes, in the end leads to the extraction of a tiny drop of new knowledge from nature. That being said, good luck to everyone scrambling for NeurIPS, you've got 10min 😋

🎉 We’re thrilled to announce the first ELLIS × Unireps Speaker Series (Reading Group associated with ELLIS PhD & Postdoc Program) session is happening soon! 🔜 🗓️ When: 29th May 2025 – CET 📍 Where: ethz.zoom.us/j/66426188160 🎙️ Who: Andrew Lampinen (Keynote) &

Why do both #ChatGPT and #Gemini UI struggle so much with rendering equations? They output beautiful, compilable latex but their UIs still manage to screw up every other equation in the app

New preprint, on prediction in vision – in mice! We've found that visual cortex, even when just viewing natural scenes, predicts *higher-level* visual features The aligns with recent developments in ML, but challenges some assumptions about early sensory cortex

THE WAY OF CODE, a project by @rickrubin in collaboration with Anthropic:

Interested in graph and topological deep learning?🍩 Join us this wednesday online for this exciting Symmetry and Geometry in Neural Representations global seminar series!🌟

Pregnancy & motherhood transform the brain in ways we're only beginning to understand🧠 We're uncovering these changes with Susana Carmona Magdalena Martínez-García PhD🦋 Emily G. Jacobs 🦋 @emilyjacobs.bsky.social Hannah Grotzinger Kaya Jordan neuromaternal thanks to CZI Science's generous support ! tinyurl.com/38jb7hm6

That's not to say things are helpless. We've gone through the wave of SAE hype and skepticism before in the context of causal representation learning. Just like with the LLM case, the hype was touched off with a set of very compelling empirical results, followed by an enormous

What is the probability of an image? What do the highest and lowest probability images look like? Do natural images lie on a low-dimensional manifold? In a new preprint with Zahra Kadkhodaie Eero Simoncelli, we develop a novel energy-based model in order to answer these questions: 🧵

The era of artificial scientific intelligence is here. As algorithms generate discoveries at scale, what role remains for human scientists?🤔 Thanks PLOS Biology for publishing my perspective Real AI @ UCSB UCSB ECE UC Santa Barbara's College of Engineering UC Santa Barbara ! journals.plos.org/plosbiology/ar…

A good language model should say “I don’t know” by reasoning about the limits of its knowledge. Our new work AbstentionBench carefully measures this overlooked skill in leading models in an open-codebase others can build on! We find frontier reasoning degrades models’ ability to

Who got time to wait for delayed generalization (grokking)? We introduce GrokAlign, a provable solution to speed up the alignment between your model and your training data resulting in faster convergence + visual probing of your DN! Ofc it uses splines :) arxiv.org/abs/2506.12284

*From superposition to sparse codes: interpretable representations in NNs* by David Klindt Nina Miolane 🦋 @ninamiolane.bsky.social Patrik Reizinger Charlie O'Neill Nice overview on the linearity of NN representations and the use of sparse coding to recover interpretable activations. arxiv.org/abs/2503.01824

Agents training in RL gym

VLMs are wild. You can glue a vision model to an LLM?! What's up? The key is that both models learn universal representations. Just training a linear projection, is enough! But image embeddings align only with LATE layer language activations as shown by LLM SAEs, causing issues

Welcome again, awesome to have you here for summer at Cold Spring Harbor Laboratory

My thesis is now online! umontreal.scholaris.ca/items/f8670d1c… This is more than just a list of publications. I invested a lot of time and passion writing this thesis in hope that it will make for an interesting read. Here's a summary of what you'll find in it.

plt.bar(*np.unique(x, return_counts=True)) –a match made in heaven 😍