Have you heard of task-, domain- and class-incremental learning? The article describing these three continual learning scenarios is now out in Nature Machine Intelligence. It’s been a while since the preprint, and there are exciting additions! 👇 nature.com/articles/s4225…

🦟 By infusing a virtual fruit fly with artificial intelligence, HHMI | Janelia and Google DeepMind scientists have created a computerized insect that can walk and fly just like the real thing. 🤖 🔗janelia.org/news/artificia…

🔍How can we design neural networks that take neural network parameters as input? 🧪Our #ICLR2024 oral on "Graph Neural Networks for Learning Equivariant Representations of Neural Networks" answers this question! 📜: arxiv.org/abs/2403.12143 💻: github.com/mkofinas/neura… 🧵 [1/9]

New preprint! Sina Tafazoli shows the brain builds complex tasks by compositionally combining simpler sub-task representations. By dynamically reusing neural subspaces for sensory inputs and motor actions, the brain can flexibly perform multiple tasks. 🧵 biorxiv.org/content/10.110…

Anxiety and depression might be related to increased, not decreased, exploration behaviour. What could be possible mechanisms underlying this? Beyond thrilled to share my very first first-author preprint with Toby Wise Quentin Huys and Helmholtz Institute for Human-Centered AI. osf.io/preprints/psya… 🧵⬇️ [1/n]

This is a fascinating result! CaMKII is needed for short-term but not long-term memory. This is reminiscent of studies from decades ago showing that the amnestic effects of protein synthesis inhibitors are sometimes transient: rodents eventually recover their memories.

𝗙𝗶𝗻𝗮𝗹𝗹𝘆 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗲𝗱 𝗶𝗻 Cell ! How does the brain processes proprioceptive information?🧠 We use a task-driven modeling approach by combining biomechanics 🦾 and neural network modeling💻 to investigate different hypotheses. 1/N cell.com/cell/pdf/S0092…

Finally a tweeprint on our recent preprint presenting Vector-HaSH ! Vector-HaSH extends MESH to unify two important and seemingly independent roles of hippocampus: Spatial Mapping and Episodic Memory ! Brief talk: youtube.com/watch?v=P4kHqk… Preprint: biorxiv.org/content/10.110… 1/n

Happy to share a new preprint - the culmination of the past three years of postdoc with Blake Richards and Adrien Peyrache 🪐😴⚡️, and my first real foray into #NeuroAI as a tool to study the sleeping brain: biorxiv.org/content/10.110… (1/🧵) c’est parti!

Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics biorxiv.org/cgi/content/sh… #biorxiv_neursci

Happy to see my thesis work out today in nature! nature.com/articles/s4158… Josh Merel Jesse Marshall lhq Yuval Tassa Ugne Klibaite Amanda Gellis Greg Wayne Matt Botvinick @BOlveczky Google DeepMind Video credit to the talented Kristian J Herrera

Happy to share our paper accepted to ICML Conference where we identify a reason (of possibly many) why most ML applications don't seem to benefit from curricula TLDR; overparameterisation-while helping in optimisation-is eroding the benefits of curricula 📄arxiv.org/abs/2406.01589 1/5

Excited to share our new work on dynamics in compositional RL with curricula accepted ICML Conference 2024! 📄 [arxiv.org/abs/2402.18361] 👯 Team = me, Stefano Sarao Mannelli Andrew Saxe TL;DR We propose a simple compositional RL model amenable to exact learning dynamics analysis which can account

Beyond Euclid: An Illustrated Guide to Modern Machine Learning with Geometric, Topological, and Algebraic Structures New review out with Nina Miolane 🦋 @ninamiolane.bsky.social, co-led by myself, Johan Mathe & Mathilde Papillon🦋 mathildepapillon .bsky .social Check out our visual taxonomies of non-Euclidean structure in data & ML models

Why is reality so slow? Why can we only have one thought at a time? Why do we need so many neurons? Will Elon Musk's Neuralink really speed up his cognition? For answers and more questions check out our new review: "The Unbearable Slowness of Being". arxiv.org/abs/2408.10234

How does sensorimotor (S1/M1) cortex support adaptive motor control?  Come find out in our latest preprint, which spans the development of a full adult forelimb model + physics simulations, neural-modeling for control, complex 🐭behavior 🕹️, large-scale imaging, and of course

Stop by poster 431.07 Board Z2 at 10a this morning at #SfN24 for “Identifying and modeling brain-wide neural activity structure in larval zebrafish,” presented by Yu Duan from Kanaka Rajan’s lab. Work in collab w/ Matt Perich & Karl Deisseroth. #KempnerInstitute Society for Neuroscience (SfN)

we've finally updated our 2⃣ year old preprint on action chunking as (conditional) policy compression! this was a lesson in perseverance and not giving up on your ideas...work with Sam Gershman & Ann_Huang ✨ (1/3) osf.io/preprints/psya…

Our new #NAACL2025 paper with Kyunghyun Cho, led by artem kirsanov, shows how different prompting strategies shape the internal representations of LLMs, through an interpretability approach rooted in statistical physics and geometry. Despite achieving similar task performance, these

New in the Deeper Learning blog: Kempner researchers show how VLMs speak the same semantic language across images and text. bit.ly/KempnerVLM by Isabel Papadimitriou, Chloe H. Su, Thomas Fel, Stephanie Gil, and Sham Kakade #AI #ML #VLMs #SAEs