How do you "see" with electric eyes? How does collective behavior emerge from individual interactions? Nocturnal weakly electric fish evolved to do this, but studying naturalistic social behavior is very hard. Our solution? Virtual 'fish' 🤖🐟⚡ 📄 arxiv.org/abs/2511.08436

Both brains and RNNs can re-use components of computation across similar tasks or contexts. But what exactly are those “shared components”? How can they be used to solve several tasks? We address these questions in a new preprint with @avm.bsky.social! Link: www.biorxiv.org/content/10.6...

Interpretable compositional computation with recurrent neural networks

Flexible cognition utilizes reusable components to enable rapid adaptation of behavior to different contexts or tasks. Analysis of artificial neural networks trained on multiple tasks suggested that t...

biorxiv.org

How can the brain learn the hidden hierarchical structure from high dimensional data? In our latest work, we use synthetic datasets to analyze two classes of bio-plausible learning rules: variants of Direct Feedback Alignment, and local self-supervised learning. We find only the latter succeeds.

Why did RNNs fail to learn long-term dependencies? What if we added one more modification, maybe now? It turns out we can give a pretty broad, analytical answer! See the attached paper for a rigorous treatment using centre manifolds, low-rank RNNs, and dynamical systems theory! go.aps.org/4fXWEeF

Ghost Mechanism: An Analytical Model of Abrupt Learning in Recurrent Networks

This study establishes the ghost mechanism as an underlying mechanism for abrupt learning, whereby the recurrent neural network develops ghost points---transient dynamical bottlenecks---and identifies...

go.aps.org

Physical Review X@physrevx.aps.org · last mo.

Why do neural networks undergo abrupt learning after training on memory tasks? New research traces this phenomenon to a “ghosting” mechanism that locks the system into a no-learning zone and proposes two practical fixes to address these pitfalls. 📝 https://go.aps.org/4fXWEeF

Can we match self-supervised backpropagation using local learning rules? We show it is possible in our new paper accepted by ICML. We achieve: 1. theoretical equivalence to BP in a controlled setup 2. new SOTA for local learning across image datasets 3. same performance as BP on multiple datasets

This is cool and makes a lot of sense. Reminds me of the theory of neutral networks in evolutionary theory, where networks of neutral genotype changes enable populations to traverse the fitness landscape without getting stuck in local minima

Flavio Martinelli@flavioh.bsky.social · 2mo ago

NEW PAPER. Why do larger networks train better? "Because they contain more candidate *sub*networks that can learn the task" → lottery tickets This popular explanation uses an appealing but misleading metaphor🧵 We propose an intuitive alternative grounded in theory: escape dimensions

Ever heard of the lottery ticket hypothesis? Our new paper shows that lottery tickets are not a useful metaphor to explain the success of overparameterized neural networks - and suggests an alternative metaphor: escape dimensions

Flavio Martinelli@flavioh.bsky.social · 2mo ago

NEW PAPER. Why do larger networks train better? "Because they contain more candidate *sub*networks that can learn the task" → lottery tickets This popular explanation uses an appealing but misleading metaphor🧵 We propose an intuitive alternative grounded in theory: escape dimensions

NEW PAPER. Why do larger networks train better? "Because they contain more candidate *sub*networks that can learn the task" → lottery tickets This popular explanation uses an appealing but misleading metaphor🧵 We propose an intuitive alternative grounded in theory: escape dimensions

🤖📊 NEW in the Deeper Learning blog: @annhuang42.bsky.social & @kanakarajanphd.bsky.social break down their recent work examining how #RNNs solve the same task in different ways, and why that matters. Joint work with @satpreetsingh.bsky.social & @flavioh.bsky.social bit.ly/4kj4fVd #NeuroAI

Measuring and Controlling Solution Degeneracy Across Task-Trained Recurrent Neural Networks - Kempner Institute

Despite reaching equal performance success when trained on the same task, artificial neural networks can develop dramatically different internal solutions, much like different students solving the sam...

bit.ly

🧵Excited to present our latest work at #Neurips25! Together with @avm.bsky.social, we discover 𝐜𝐡𝐚𝐧𝐧𝐞𝐥𝐬 𝐭𝐨 𝐢𝐧𝐟𝐢𝐧𝐢𝐭𝐲: regions in neural networks loss landscapes where parameters diverge to infinity (in regression settings!) We find that MLPs in these channels can take derivatives and compute GLUs 🤯

📍Excited to share that our paper was selected as a Spotlight at #NeurIPS2025! arxiv.org/pdf/2410.03972 It started from a question I kept running into: When do RNNs trained on the same task converge/diverge in their solutions? 🧵⬇️

Exciting news for #drosophila #connectomics and #neuroscience enthusiasts: the Drosophila male central nervous system connectome is now live for exploration. Find out more at the landing page hosted by our Janelia FlyEM collaborators www.janelia.org/project-team....

Male CNS Connectome

A team of researchers has unveiled the complete connectome of a male fruit fly central nervous system —a seamless map of all the neurons in the brain and nerve cord of a single male fruit fly and the ...

janelia.org

Lab members are at the Bernstein conference @bernsteinneuro.bsky.social with 9 posters! Here’s the list: TUESDAY 16:30 – 18:00 P1 62 “Measuring and controlling solution degeneracy across task-trained recurrent neural networks” by @flavioh.bsky.social

To our fellow researchers at Harvard and elsewhere. 🧪🧠 I have funds for visiting PhDs or postdocs at TU in Vienna. For short stay or full PhD email me. For professors, check for instance, this tenure track opening or ask in private for options informatics.tuwien.ac.at/news/2909