Jesseba Fernando

Already counting down till next year ⏳

Packed room at our satellite!!

On wednesday, @jesseba.bsky.social @marilyngatica.bsky.social with @nicopedre.bsky.social organize Cobrains kgatica.github.io/CCS2025-CoBr....

and then we double down with @jesseba.bsky.social talking about adaptive networks in a learn-relearning generalization task in rodents vs artificial networks (Biological networks #2, 2/9, with @scarpino.bsky.social).

The forgotten half of scientific thinking | PNAS The forgotten half of scientific thinking

“The idea that taking walks, reading things unrelated to your research, and hanging out with strangers in a campus pub should be considered part of the serious process of thinking, but might well meet with skepticism in practice.”

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

- did not measure people's creativity performance by any reputable creativity task or measure, just measured the ESSAY content (after giving people different instructions)

- did not measure people's overall MEMORY ABILITY

- false on "brain activity," DIFFERENT activity is not "weaker"

it’s even more demoralizing when the history you’re repeating is history you were around for the first time

No Kings On June 14—Flag Day—Donald Trump wants tanks in the street and a made-for-TV display of dominance for his birthday. A spectacle meant to look like strength. But real power isn’t staged in Washington. ...

www.nokings.org

Without me, Trump would have lost the election, Dems would control the House and the Republicans would be 51-49 in the Senate.

This must have been what it felt like at the end of the first triumvirate. This is so cool, I love living through history.

Super, super excited to be here at the sixth annual conference on the Mathematics of Neuroscience and AI!

Hot takes/kind reminders

a) “the bitter lesson”’s promise of scale did not deliver wrt AI reasoning & higher cog functions—O3 hallucinates 2x more than O1

b) prediction is not understanding, hallucinating is worse

c) gen AI could help neuroAI but following latest AI trends is not a paradigm shift

This was so much fun!!!

Brainstem sensing of multiple body signals during food consumption Studies of body-to-brain communication often examine one stimulus or organ at a time, yet the brain must integrate many body signals during behavior. For example, food consumption generates diverse or...

I’m so excited to share our preprint on how brainstem neurons sense and integrate multiple body signals during food consumption. We imaged 1000s of neurons across the lateral parabrachial nucleus (LPBN) in behaving mice. www.biorxiv.org/content/10.1...

m.youtube.com/watch?time_c...

I guess it’s homegrown crim of me to publicize my thoughts on Columbia rolling over so easily

Coinciding with it's "No" letter Harvard has refashioned its university homepage into essentially an advertisement for the social benefits of university research. This is the current front page.

We are rallying at noon at Cambridge Common for the leadership of the Harvard Corporation to stand up and stand with us as we defend higher education and democracy in the United States from an unprecedented attack.

Over 800 faculty have urged Harvard's leadership to "legally contest ... unlawful demands that threaten academic freedom and university self-governance." The Cambridge City Council has called on it to "stand up in defense of the values that are fundamental to both the University and our democracy."

MARCHING EVERYWHERE: The Largest Coordinated Protests in US History. A chart with detailed information about 8 different massive days of coordinated protest in U.S. history. Apologies that there is not room here to copy all the data. But the upshot is that today's Hands Off mobilization stands out for the number of locations where protests took place.

A map of the United States with a sea of red dots fully covering the blue background except in some of the most rural areas of the country. Beneath the map are the words "Map and List of Nationwide 'Hands Off' Protest Locations on April 5

For the data nerds out there, here's a chart of the largest coordinated protests in U.S. history from my 2018 book HOW TO READ A PROTEST

Today's "Hands Off" protests, in 1000+ locations around the country, rank as one of the very largest decentralized days of protests EVER.

Hands Off! Donald Trump and Elon Musk think this country belongs to them. They're taking everything they can get their hands on and daring the world to stop them. On Saturday, April 5th, we're taking to the stre...

If you’re upset about the chaos and corruption Trump is inflicting on your country—whether your focus is the nihilistic dismantling of government, the economy, our foreign relations, or his threats to the rule of law—please join the nationwide marches on April 5. Learn more here 👉

handsoff2025.com

If you are a faculty member at a US university, please consider signing this excellent petition asking our presidents, chancellors, etc to work together to defend our institutions from the direct attacks by Trump administration: sites.google.com/view/we-must...

I think some in the Senate have convinced themselves that the backlash to folding will be limited to “the left” and therefore ignorable.

Those people are deeply and profoundly misreading the moment and the stakes.

Virtually every swing district House Dem walked the plank to vote NO for a reason.

Transformer Dynamics: A neuroscientific approach to interpretability of large language models As artificial intelligence models have exploded in scale and capability, understanding of their internal mechanisms remains a critical challenge. Inspired by the success of dynamical systems approache...

12/ Neuroscientists have been doing MechInterp on the brain for many decades now, and while access to the underlying system has been shallow for the brain vs. NNs, the theory is advanced and (we argue) could still be useful for understanding AI.

The paper: arxiv.org/abs/2502.12131

(End of thread)

11/ Finally, we asked: Does the RS behave like a dynamical system?

To test this, we teleported RS vectors to different positions in PCA space and observed their evolution.

Result? Early layers exhibit attractor-like behavior—pushing activations back toward their “natural” trajectory.

10/ To visualize global RS behavior, we trained a Compressing Autoencoder (CAE)—squashing activations into 2D.

The CAE finds low-dimensional population dynamics, which showed that early layers are harder to reconstruct, while later layers become lower-dimensional and more structured.

9/ Looking at individual unit dynamics, we found rotational trajectories in activation space!

Some units spin around a fixed point, others spiral outward. On average, RS units circle ~10 times over 64 layers.

8/ And here’s something even weirder: mutual information drops in early layers, then slowly climbs back up. Why? We’re not sure—perhaps early layers are quite predictable from one to the next, but in a nonlinear way...

7/ Zooming out: the cosine similarity of the RS vector between layers increases.

Translation: The model’s internal representations become more aligned as we move deeper through the model. But they also accelerate!

6/ Second surprise: RS activations are highly correlated layer to layer.

A given unit’s activation is highly correlated across layers—especially within a layer (attention to MLP) vs. across layers.

5/ First surprise: RS activations increase in density over layers.

Early layers are sparse, later layers are dense. Why? There’s no a priori reason for this—blocks could just as easily write balancing negative values. But they don’t.