congrats Jure, this is incredibly cool data!

Thanks for the interest!! I’m curious what you mean about the UMAP, was it on the image responses?

Agreed with the point that manifolds are probably different between contexts… I guess the question is whether the manifolds are constrained by physical circuits (a la Langdon & Engel) or simply defined by tasks. I also don’t really have a horse in the race!

Yes, exactly. I agree with you that most claims of low-D neural activity are probably high-D by our definition. But, the underlying dynamics can still be low-D.

Not necessarily! See our conjecture with Valentin: given a neural nonlinearity (p-times diff’able), you can get a spectrum with slope (alpha) close to 1 from low-dim “pre-activations” (inputs of dimension d)

congrats!!

"Low-d neural activity" would probably look high-D if you plot the spectrum on a log-log scale. @engeltatiana.bsky.social made a similar point recently on @braininspired.bsky.social. Also agrees with our claim with @bio-emergent.bsky.social that low-D and high-D can be two sides of the same coin!

The power laws don't necessarily pose a contradiction to low-D dynamics!

see: openreview.net/forum?id=cGk...

See our paper with @bio-emergent.bsky.social et al openreview.net/forum?id=cGk...

My (speculative) interpretation is that spont. activity in mouse cortex is a nonlinear readout of low-D global variables. I think activity lives in a low-D manifold where the axes correspond to arousal/internal state/neuromodulation levels, and neural nonlinearities project it into high-D in cortex.

I don't think it's crazy to think there are truly low-D manifolds! We see high-D power-law eigenspectrum in mouse cortex, both in image responses and spontaneous activity (no low-D task!). But, we find that spontaneous activity has MUCH lower "pre-activation" dimensionality than image responses.

Example from our NeurIPS talk with @bio-emergent.bsky.social :

In an RNN with low-D *pre-activation* dynamics, an activation function produces high-D *post-activations*.

A downstream neuron could only read out boring sinusoids from low-D state, but can read any function from high-D activity!

Low-d dynamics are useful - they are noise robust and easier to control.

BUT you also get a major benefit from curling up a low-D state manifold in a high-D embedding space: you can then linearly read out arbitrary functions of state!

thanks!! hmm maybe a lab retreat 👀

thank you Shuqi!! was so nice to meet you through this work :)

the paper: www.biorxiv.org/content/10.1...

Thanks to authors @shuqiw.bsky.social, Yixiao, @carandinilab.net and @kenneth-harris.bsky.social! And to our photographer @flavioh.bsky.social :)

Best part about being a scientist is the people I get to work with. Valentin (@bio-emergent.bsky.social) and I got to give a talk at NeurIPS, bridging a gap between low- and high-dim perspectives of the brain. Thankfully, the audience was (somewhat) more awake than the San Diego desert 🏜️

Tomorrow at #NeurIPS2025! Oral at 10 am in UL Ballroom 20D and poster #2016 at 11 am. @haydari.bsky.social and I are looking forward to hearing your thoughts.

congrats!!

congrats Tom, Montreal is lovely I’m sure you’ll enjoy it!

🎉 "High-dimensional neuronal activity from low-dimensional latent dynamics: a solvable model" will be presented as an oral at #NeurIPS2025 🎉

Feeling very grateful that reviewers and chairs appreciated concise mathematical explanations, in this age of big models.

www.biorxiv.org/content/10.1...
1/2

Null and Noteworthy: Downstream brain areas read visual cortex signals en masse in mice The finding contradicts a theory that the regions prioritize neurons that are adept at identifying specific stimuli. Plus, a response to a study that questioned immune memory in astrocytes.

Some neurons respond more strongly to visual stimuli than others. But the rest of the brain doesn’t prioritize signals from those neurons, according to a new study. Read more in this month’s Null and Noteworthy.

By @ldattaro.bsky.social

#neuroskyence

www.thetransmitter.org/null-and-not...

Introducing our new favorite stimulus. A few minutes are enough to map the visual preferences of thousands of neurons.

Mapping the visual cortex with Zebra noise and wavelets
www.biorxiv.org/content/10.1...

By Sophie Skriabine and Max Shinn, with Samuel Picard and
@kenneth-harris.bsky.social

New by Agnès Landemard (@agnesland.bsky.social‬) & co

Brainwide blood volume reflects opposing neural populations

Brainwide fluctuations in blood volume arise from two populations with opposite relation to brain state and distinct relationships to blood supply

www.biorxiv.org/content/10.1...

Our new preprint 👀

High-dimensional neuronal activity from low-dimensional latent dynamics: a solvable model Computation in recurrent networks of neurons has been hypothesized to occur at the level of low-dimensional latent dynamics, both in artificial systems and in the brain. This hypothesis seems at odds ...

The firing of neural populations is high-dim even if their subthreshold activity is low-dim! This work by @bio-emergent.bsky.social and @haydari.bsky.social shows how, with a solvable model, a data analysis technique, and data from mouse visual cortex: www.biorxiv.org/content/10.1...

Excited to share my PhD paper! In it, we use targeted 2-photon optogenetic stimulation to determine how V1 activity is read-out in a detection task. We found that network influence, not visual coding properties, predicted the impact of ensembles on behavior - contradicting our expectations (1/5).

So cool! Congrats :)