Shahab Bakhtiari

Vectorized instructive signals in cortical dendrites - Nature Mice learning a neurofeedback brain–computer interface task show neuron-specific teaching signals in cortical dendrites, consistent with a vectorized solution for credit assignment in the brain.

This paper on how the brain may do gradient descent is very cool: www.nature.com/articles/s41...

Pace of ecology drives the tempo of visual perception across the animal kingdom Nature Ecology & Evolution - Using phylogenetic comparative methods across 237 species from disparate phyla, the authors show that species with fast-paced ecologies have higher temporal...

Our new paper is now out showing how time perception in animals is linked to their ecology. Using data from 237 species we show temporal perception is faster in species that fly and pursuit predators www.nature.com/articles/s41... 🌐

This study is super cool (connecting ecology and perception), that suggest some aspects of animal's perception (temporal precision) is shaped by their environment (which somehow resonates w our proposal on internal foraging perspectives on perceptual selection www.sciencedirect.com/science/arti...)

This is a critical methodological point about the Platonic Representation Hypothesis paper.

I mistakenly thought the PRH paper used CKA as its main similarity metric.

Another motivation for thinking more deeply about metrics of similarity and alignment.

Though it’s quite interesting that this subtle methodological detail turned out to be so important in the main message.

Thanks for the correction!

Yes, in the main text, your paper mainly relied on local similarity.

Now actually I rememberer, my first reaction reading your paper was why CKA results weren’t used in the main text.

Humans and neural networks show similar patterns of transfer and interference during continual learning - Nature Human Behaviour When learning new tasks, both humans and artificial neural networks face a trade-off between reusing prior knowledge to learn faster and avoiding the disruption of earlier learning. This study shows t...

What you describe sounds like the "lumpers vs. splitters" in this paper from @summerfieldlab.bsky.social lab: lumpers generalize more/retain less, and splitters generalize less/forget more. They gave a nice explanation based on rich vs lazy training regimes in ANNs.

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

Looking forward to reading the promised post on continual learning, @lampinen.bsky.social :)

Cool! This is generative inference’s prediction for human perception in this illusion: the squares are no longer squares!
Try it for yourself here:
huggingface.co/spaces/ttoos...

Memorization vs. generalization in deep learning: implicit biases, benign overfitting, and more Or: how I learned to stop worrying and love the memorization

What is the relationship between memorization and generalization in AI? Is there a fundamental tradeoff? In infinitefaculty.substack.com/p/memorizati... I’ve reviewed some of the evolving perspectives on memorization & generalization in machine learning, from classic perspectives through LLMs.

Revisiting the Platonic Representation Hypothesis: An Aristotelian View The Platonic Representation Hypothesis suggests that representations from neural networks are converging to a common statistical model of reality. We show that the existing metrics used to measure rep...

Another good reason for being cautious with representational similarity analysis: arxiv.org/abs/2602.14486

The famous Platonic Representation Hypothesis was largely driven by CKA’s bias.

But the hypothesis still holds for shared local relationships.

The revised version of our paper on the impact of top-down feedback is now out @elife.bsky.social:

doi.org/10.7554/eLif...

tl;dr: we show that using human-brain-like feedback/anatomy in a deep RNN leads to human-like visual biases!

This work was led by @tmshbr.bsky.social

#NeuroAI 🧠📈 🧪

Excited to launch Principia, a nonprofit research organisation at the intersection of deep learning theory and AI safety.

Our goal is to develop theory for modern machine learning systems that can help us understand complex network behaviors, including those critical for AI safety and alignment.

1

Thrilled to finally share this work! 🧠🔊

Using a new reinforcement-free task we show mice (like humans) extract abstract structure from sound (unsupervised) & dCA1 is causally required by building factorised, orthogonal subspaces of abstract rules.

Led by Dammy Onih!
www.biorxiv.org/content/10.6...

I don’t think AI’s success in coding will automatically translate to other fields. That level of performance only works where the output is as easily verifiable as code; and not many domains fit that bill. 2/2

Something big is happening in AI — and most people will be blindsided | Fortune It’s not like a light switch... more like the moment you realize the water has been rising around you and is now at your chest.

"The experience that tech workers have had over the past year, of watching AI go from “helpful tool” to “does my job better than I do”, is the experience everyone else is about to have. Law, finance, medicine, accounting, …"

I’m not sure … 1/2

fortune.com/2026/02/11/s...

… especially whenever controversies around representational similarity resurface.

You’re comparing two fields at very different stages of theoretical maturity. Neuroscience (and NeuroAI) are still largely pre-theoretic. I often return to Hasok Chang’s Inventing Temperature as a parallel for where we actually stand in theoretical neuroscience, …

Definitely not enough.

My bet is on the ecological relevance of training data and temporal prediction as the core objective.

Architecture is difficult to constrain, given that ANNs and brains rely on substantially different functional mechanisms.

That’s just my view, though; I could be wrong.

Exactly my point. The emerging view seems to be that, assuming equal trainability (big assumption though), the architectures may not play as big of a role as the training objective and data.

The problem is the training objective and lack of recurrence, not the 50-layer architecture.

Also see @mschrimpf.bsky.social thread here: bsky.app/profile/msch...

My two cents: as with any discipline, we use tools to probe phenomena scientifically while simultaneously striving to understand those tools better

An iterative process that sees the works like the one above within its paradigm not standing outside of it.

I encourage everyone to read the paper and rather than relying on social media impressions.

It’s an excellent paper speaking about a specific category of work in NeuroAI that is still hard to generalize to the whole field.

You mean a framework that could build models of artificial stimuli would be sufficient?

In praise of artifice - Nature Neuroscience Nature Neuroscience - In praise of artifice

Great point.

I’d use classical stimuli for testing out-of-distribution generalization instead of model development.

This would actually be the exact opposite of what was proposed in "In praise of artifice"

www.nature.com/articles/nn1...

Our paper is out in @natneuro.nature.com!

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

We develop a geometric theory of how neural populations support generalization across many tasks.

@zuckermanbrain.bsky.social
@flatironinstitute.org
@kempnerinstitute.bsky.social

1/14

I agree. My main point is that decompositionality (whether or not it supports modularity) is baked into classical stimuli a priori. These stimuli then act as an inductive bias in models developed to capture the resulting neural or behavioral responses.

n other words, classic stimuli rely on a strong notion of modularity that may or may not hold for naturalistic stimuli, where visual features are inherently intermixed.