Anirudh GJ

Three panel thing. In the left panel we use error bars. In the second, we take statistical significance as the biggest number but still have error bars. In LLM science, we just have the biggest number

What if we did a single run and declared victory

Very excited to release a new blog post that formalizes what it means for data to be compositional, and shows how compositionality can exist at multiple scales. Early days, but I think there may be significant implications for AI. Check it out! ericelmoznino.github.io/blog/2025/08...

📰 I really enjoyed writing this article with @thetransmitter.bsky.social! In it, I summarize parts of our recent perspective article on neural manifolds (www.nature.com/articles/s41...), with a focus on highlighting just a few cool insights into the brain we've already seen at the population level.

YouTube video by Gerstner Lab From Spikes To Rates

Is it possible to go from spikes to rates without averaging?

We show how to exactly map recurrent spiking networks into recurrent rate networks, with the same number of neurons. No temporal or spatial averaging needed!

Presented at Gatsby Neural Dynamics Workshop, London.

I wonder, where would be a good place to do modeling and chat with many people that study different species or do comparative studies? (asking for a friend)

A summary figure for a NeurIPS competition where AI agents compete with mice in a visual foraging task.

Mice learn these tasks and are robust to perturbations like fog. Now, we invite you all to make AI agents to beat mice.

We present our #NeurIPS competition. You can learn about it here: robustforaging.github.io (7/n)

Simple low-dimensional computations explain variability in neuronal activity Our understanding of neural computation is founded on the assumption that neurons fire in response to a linear summation of inputs. Yet experiments demonstrate that some neurons are capable of complex...

This paper carefully examines how well simple units capture neural data.

To quote someone from my lab (they can take credit if they want):

Def not news to those of us who use [ANN] models, but a good counter argument to the "but neurons are more complicated" crowd.

arxiv.org/abs/2504.08637

🧠📈 🧪

Working memory control dynamics follow principles of spatial computing - Nature Communications It is unclear how cognitive computations are performed on sensory information. Here, neural evidence from working memory tasks suggests that the physical dimensions of cortical networks are used to up...

"These findings validate core predictions of Spatial Computing by showing that oscillatory dynamics not only gate information in time but also shape where in the cortex cognitive content is represented."
More on Spatial Computing:
doi.org/10.1038/s414...

Structure of activity in multiregion recurrent neural networks | PNAS Neural circuits comprise multiple interconnected regions, each with complex dynamics. The interplay between local and global activity is thought to...

(1/23) In addition to the new Lady Gaga album "Mayhem," my paper with Manuel Beiran, "Structure of activity in multiregion recurrent neural networks," has been published today.

PNAS link: www.pnas.org/doi/10.1073/...

(see dclark.io for PDF)

An explainer thread...

Universality and diversity in human song Songs exhibit universal patterns across cultures.

Music is universal. It varies more within than between societies and can be described by a few key dimensions. That’s because brains operate by using the raw materials of music: oscillations (brainwaves).
www.science.org/doi/10.1126/...
#neuroscience

1/N
How do neural dynamics in motor cortex interact with those in subcortical networks to flexibly control movement? I’m beyond thrilled to share our work on this problem, led by Eric Kirk @eric-kirk.bsky.social with help from Kangjia Cai!
www.biorxiv.org/content/10.1...

Thrilled to announce I'll be starting my own neuro-theory lab, as an Assistant Professor at @yaleneuro.bsky.social @wutsaiyale.bsky.social this Fall!

My group will study offline learning in the sleeping brain: how neural activity self-organizes during sleep and the computations it performs. 🧵

screenshot of biorxiv paper titled "Neuromorphic hierarchical modular reservoirs", authors Filip Milisav, Andrea I Luppi, Laura E Suárez, Guillaume Lajoie, Bratislav Misic

aside from this being a v cool paper I also want to congratulate the authors on the incredible SNR achieved in the title via a complete absence of filler words

Neuromorphic hierarchical modular reservoirs
www.biorxiv.org/content/10.1...

New preprint! 🧠🤖

How do we build neural decoders that are:
⚡️ fast enough for real-time use
🎯 accurate across diverse tasks
🌍 generalizable to new sessions, subjects, and even species?

We present POSSM, a hybrid SSM architecture that optimizes for all three of these axes!

🧵1/7

Curious about the history of the manifold/trajectory view of neural activity.

My own first exposure was Gilles Laurent's chapter in "21 Problems in Systems Neuroscience", where he cites odor trajectories in locust AL (2005). This was v inspiring as a biophysics student studying dynamical systems...

I think the biological evidence points to this not being the case. We can see instances where synapses literally undergo a form of reverse plasticity, e.g. see here: www.cell.com/trends/cogni...

I think it cannot be assumed that we never wipe memories from our brains completely!

How a neuroscientist solved the mystery of his own long COVID

How a neuroscientist solved the mystery of his own
#LongCovid and lead to a new scientific discovery. Inspiring story.
Thank you for sharing your journey @jeffmyau.bsky.social
www.youcanknowthings.com/how-one-neur...

Thalamus: a brain-inspired algorithm for biologically-plausible continual learning and disentangled representations Animals thrive in a constantly changing environment and leverage the temporal structure to learn well-factorized causal representations. In contrast, traditional neural networks suffer from forgetting...

I love how this paper uses cortico-thalamic interactions(context switching) for continual learning.
arxiv.org/abs/2205.11713

RL promises "systems that can adapt to their environment". However, no RL system that I know of actually fulfill anything close to this goal, and, furthermore, I'd argue that all the current RL methodologies are actively hostile to this goal. Prove me wrong.

Comparing cooperative geometric puzzle solving in ants versus humans | PNAS Biological ensembles use collective intelligence to tackle challenges together, but suboptimal coordination can undermine the effectiveness of grou...

This is also one of the reasons why autonomous vehicles will eventually be much better than humans.

www.pnas.org/doi/10.1073/...

Flexibility of intrinsic neural timescales during distinct behavioral states - Communications Biology Calcium imaging of spontaneously behaving mice show increased intrinsic neural timescales during behavior. The behavioral state of mice can be predicted from the topography of timescales of the cortex...

Flexibility of intrinsic neural timescales during distinct behavioral states
www.nature.com/articles/s42...
#neuroscience

No More Adam: Learning Rate Scaling at Initialization is All You Need In this work, we question the necessity of adaptive gradient methods for training deep neural networks. SGD-SaI is a simple yet effective enhancement to stochastic gradient descent with momentum (SGDM...

This paper looks interesting - it argues that you don’t need adaptive systems like Adam to get good gradient-based training, instead you can just set a learning rate for different groups of units based on initialization:

arxiv.org/abs/2412.11768

#MLSky #NeuroAI

Single cortical neurons as deep artificial neural networks Utilizing recent advances in machine learning, we introduce a systematic approach to characterize neurons’ input/output (I/O) mapping complexity. Deep…

There is also this one: www.sciencedirect.com/science/arti...

sinthlab EoY social! I'm grateful everyday that I get to work with such a kind and intelligent group of individuals.

@mattperich.bsky.social @oliviercodol.bsky.social @anirudhgj.bsky.social

Neural Attention Memory Models are evolved to optimize the performance of Transformers by actively pruning the KV cache memory. Surprisingly, we find that NAMMs are able to zero-shot transfer its performance gains across architectures, input modalities and even task domains! arxiv.org/abs/2410.13166

Thrilled to share a new preprint exploring the spatial organization of multisensory convergence in the mouse isocortex! 🧠🎉 Even more special as it builds on work I started during my undergrad, a lifetime ago 👨‍🦳

Check it out here: www.biorxiv.org/content/10.1...

Cerebellar-driven cortical dynamics enable task acquisition, switching and consolidation To drive behavior, the cortex must bridge sensory cues with future outcomes. However, the principles by which cortical networks learn such sensory-behavioural transformations remain largely elusive. H...

Indeed! This is in line with our cerebellar-cortical models, in which we show that cerebellar-feedback greatly alleviates the need for plasticity/gradients in the cortex! This in turn explains quite a bit of obs: doi.org/10.1101/2022... (out soon!)

Feedback control guides credit assignment in recurrent neural networks How do brain circuits learn to generate behaviour? While significant strides have been made in understanding learning in artificial neural networks, applying this knowledge to biological networks...

This paper from Claudia Clopath's lab is very exciting for me.

I've been musing for a while now that the challenge of approximating gradient descent is probably made way easier in the brain by feedback signals for control.

Here they show it!!!

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

🧠📈 #NeuroAI 🧪

Neuronal sequences in population bursts encode information in human cortex - Nature The temporal order of neuronal firing within bursts of population spiking in the human anterior temporal lobe is dependent on the category as well as the identity of the individual stimulus, and this ...

This study shows that spike sequences carry information beyond what rates and latency to first spike do:

www.nature.com/articles/s41...

My reactions:

1) Cool to see this in humans.
2) Are people still surprised that spike times carry information beyond rates/first-spike latency?!?!?!

🧠📈 🧪

I’m speechless. Truly an art and despite being only four minutes long, I learned a lot!