Gatsby Computational Neuroscience Unit

UCL – University College London UCL is consistently ranked as one of the top ten universities in the world (QS World University Rankings 2010-2022) and is No.2 in the UK for research power (Research Excellence Framework 2021).

JOB: We're recruiting for a 2-year post doc fellowship based at @gatsbyucl.bsky.social working with Prof Arthur Gretton to create new statistically efficient causal and interaction models with the aim of elucidating cellular interactions
⏰closing 27 August
👉 details: www.ucl.ac.uk/work-at-ucl/...

Dopamine encodes deep network teaching signals for individual learning trajectories Longitudinal tracking of long-term learning behavior and striatal dopamine reveals that dopamine teaching signals shape individually diverse yet systematic learning trajectories, captured mathematical...

Does the brain learn by gradient descent?

It's a pleasure to share our paper at @cp-cell.bsky.social, showing how mice learning over long timescales display key hallmarks of gradient descent (GD).

The culmination of my PhD supervised by @laklab.bsky.social, @saxelab.bsky.social and Rafal Bogacz!

Aeon: an open-source platform to study the neural basis of ethological behaviours over naturalistic timescales Ethological behaviours are a powerful tool for neuroscience since they leverage the robust neural computations shaped by the species' evolution to study the neural basis of high-level cognitive functi...

New from @sainsburywellcome.bsky.social, @gatsbyucl.bsky.social, NeuroGEARS & DataJoint:

Aeon - an open-source platform combining long-term neural recording with naturalistic behavioural tracking in enriched habitats.

📑 Preprint: www.biorxiv.org/content/10.1...

🎓Thrilled to share I’ve officially defended my PhD!🥳

At @gatsbyucl.bsky.social, my research explored how prior knowledge shapes neural representations.

I’m deeply grateful to my mentors, @saxelab.bsky.social and @caswell.bsky.social, my incredible collaborators, and everyone who supported me!

🥳 Congratulations to Rodrigo Carrasco-Davison on passing his PhD viva with minor corrections! 🎉

📜 Principles of Optimal Learning Control in Biological and Artificial Agents.

Come chat about this at the poster @icmlconf.bsky.social, 11:00-13:30 on Wednesday in the West Exhibition Hall #W-902!

Excited to share new work @icmlconf.bsky.social by Loek van Rossem exploring the development of computational algorithms in recurrent neural networks.

Hear it live tomorrow, Oral 1D, Tues 15 Jul West Exhibition Hall C: icml.cc/virtual/2025...

Paper: openreview.net/forum?id=3go...

(1/11)

Distributional diffusion models with scoring rules at #icml25

Fewer, larger denoising steps using distributional losses!

Wednesday 11am poster E-1910

arxiv.org/pdf/2502.02483

@vdebortoli.bsky.social
Galashov Guntupalli Zhou
@sirbayes.bsky.social
@arnauddoucet.bsky.social

Accelerated Diffusion Models via Speculative Sampling, at #icml25 !

16:30 Tuesday July 15 poster E-3012

arxiv.org/abs/2501.05370

@vdebortoli.bsky.social Galashov @arnauddoucet.bsky.social

Leena Chennuru Vankadara (@leenacvankadara.bsky.social) will also give:
- a tutorial 'Training Neural Networks at Any Scale' (Mon 14 Jul)
- and a talk at ‪WiML Symposium at ICML (Wed 16 Jul).
Don't miss out!

List of posters by/involving Gatsby Unit researchers

👋 Attending #ICML2025 next week?
Don't forget to check out work involving our researchers!

🎊 Congratulations to Kevin Huang (@kevhhuang.bsky.social‬) on passing his PhD viva with minor corrections! 🥳

🔖 Universality beyond the classical asymptotic regime

A-Level Students from Camden visiting the Advanced Microscopy Facilities at the Sainsbury Wellcome Centre.

It was a pleasure to welcome work experience students to SWC and @gatsbyucl.bsky.social this week, through the Camden STEAM initiative with @ucllifesciences.bsky.social

They explored our labs and learned about the many paths into #neuroscience

www.sainsburywellcome.org/web/content/...

How do task dynamics impact learning in networks with internal dynamics?

Excited to share our ICML Oral paper on learning dynamics in linear RNNs!
with @clementinedomine.bsky.social @mpshanahan.bsky.social and Pedro Mediano

openreview.net/forum?id=KGO...

🥳 Congratulations to Clementine Domine (@clementinedomine.bsky.social) on passing her PhD viva with minor corrections! 🎉

🔖 Balancing Learning Regimes: The Impact of Prior Knowledge on the Dynamics of Neural Representations

Dimitri Meunier, Antoine Moulin, Jakub Wornbard, Vladimir R. Kostic, Arthur Gretton
Demystifying Spectral Feature Learning for Instrumental Variable Regression
https://arxiv.org/abs/2506.10899

Schematic of the study

New research shows long-term learning is shaped by dopamine signals that act as partial reward prediction errors.

The study in mice reveals how early behavioural biases predict individual learning trajectories.

Find out more ⬇️

www.sainsburywellcome.org/web/blog/lon...

Poster from UCL Faculty of Life Sciences headed in olive green, under which are images of cyanotypes (whole image to left in blue and white, partial detail to right in blue & green on dark background) under which is a substantially blue and white panel detailing the event, speakers and background info - all available more accessibly via the online link.

Our #Artist-in-Residence, Mary Yacoob, has produced new large-scale #cyanotype prints reimagining of #scientific visual data from @ucllifesciences.bsky.social #researchers. Faculty private #exhibition details here: tinyurl.com/YacoobUCL
#artandscience #art #artwork @ucl.ac.uk @ucl-cdb.bsky.social

Last but not least of my travel updates: Courtesy of the very kind @slchau.bsky.social, I'm giving a talk at 14:30, 27 Jun
@NTU CCDS in SG on data augmentation & Gaussian universality, which strings together several works over my PhD.

If you're in SG/Lyon over the next few weeks, let me know! :)

Universality of High-Dimensional Logistic Regression and a Novel CGMT under Dependence with Applications to Data Augmentation Over the last decade, a wave of research has characterized the exact asymptotic risk of many high-dimensional models in the proportional regime. Two foundational results have driven this progress: Gau...

Meanwhile, excited to be in #Lyon for #COLT2025, with a co-first author paper (arxiv.org/abs/2502.15752) with the amazing team -- Matthew M Mallory and our advisor Morgane Austern!

Keywords: Gaussian universality, dependent data, convex Gaussian min-max theorem, data augmentation!

Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schrödinger Equation Incorporating group symmetries into neural networks has been a cornerstone of success in many AI-for-science applications. Diagonal groups of isometries, which describe the invariance under a simultan...

Sadly missing ICML due to visa :'(, but looking forward to share our ICML paper (arxiv.org/abs/2502.05318) as a poster at #BayesComp, Singapore!

Work on symmetrising neural nets for schrodinger equation in crystals, with the amazing Zhan Ni, Elif Ertekin, Peter Orbanz and Ryan P. Adams

Composite Goodness-of-fit Tests with Kernels, now out in JMLR!

www.jmlr.org/papers/v26/2...

Test if your distribution comes from ✨any✨ member of a parametric family. Comes in MMD and KSD flavours, and with code.

@oscarkey.bsky.social @fxbriol.bsky.social Tamara Fernandez

We’ve successfully automated part of the (neuro)scientific process. Now I may be out of a job. After a lot of iterations, here’s our framework for automatic discovery, built to embrace degeneracy and the realities of underconstrained modeling.

How does in-context learning emerge in attention models during gradient descent training?

Sharing our new Spotlight paper @icmlconf.bsky.social: Training Dynamics of In-Context Learning in Linear Attention
arxiv.org/abs/2501.16265

Led by Yedi Zhang with @aaditya6284.bsky.social and Peter Latham

Dimitri Meunier, Zikai Shen, Mattes Mollenhauer, Arthur Gretton, Zhu Li
Optimal Rates for Vector-Valued Spectral Regularization Learning Algorithms
https://arxiv.org/abs/2405.14778

Nonlinear Meta-Learning Can Guarantee Faster Rates Many recent theoretical works on \emph{meta-learning} aim to achieve guarantees in leveraging similar representational structures from related tasks towards simplifying a target task. The main aim of ...

🚨 New paper accepted at SIMODS! 🚨
“Nonlinear Meta-learning Can Guarantee Faster Rates”

arxiv.org/abs/2307.10870

When does meta learning work? Spoiler: generalise to new tasks by overfitting on your training tasks!

Here is why:
🧵👇

RL Perceptron: Generalization Dynamics of Policy Learning in High Dimensions A solvable model for reinforcement learning, the RL perceptron, provides a mathematical framework to analyze learning dynamics, revealing key efficiency factors and a speed-accuracy trade-off that can...

Our paper just came out in PRX!

Congrats to Nishil Patel and the rest of the team*

TL;DR: We analyse policy learning through the lens of statphys, revealing distinct scaling regimes with sharp transitions.

🔗 journals.aps.org/prx/abstract...

*Seb Lee @sebgoldt.bsky.social @saxelab.bsky.social

Mattes Mollenhauer, Nicole M\"ucke, Dimitri Meunier, Arthur Gretton: Regularized least squares learning with heavy-tailed noise is minimax optimal https://arxiv.org/abs/2505.14214 https://arxiv.org/pdf/2505.14214 https://arxiv.org/html/2505.14214

Triggered cell death in the tail of the striatum using a viral strategy. Each image is an average projection of a part of the striatum. Areas are colour coded based on their absence of neurons. Original colours are random to indicate the amount of animals that do not have cells on a particular region. The average projections then creates a mix of those original colours. Credit: Hernando Martinez Vergara

SWC researchers reveal a new way that the brain learns.

The second learning system is frequency-based and may explain how the brain forms habits and why they are so hard to break.

Full story: www.sainsburywellcome.org/web/research...

Our paper "Solve Layerwise Linear Models First to Understand Neural Dynamical Phenomena" has been accepted as a position paper to ICML 2025!
arxiv.org/abs/2502.21009
These models offer a tractable path to understanding complex neural dynamics—before diving into full nonlinearity. 1/4