I'm more and more convinced that low-dimensional manifolds in the brain are just an artifact of the experimental designs and analyses we use...
🧠📈 🧪
Dimensionality reduction may be the wrong approach to understanding neural representations. Our new paper shows that across human visual cortex, dimensionality is unbounded and scales with dataset size—we show this across nearly four orders of magnitude.
journals.plos.org/ploscompbiol...
For this paper, it’s worth considering that structure in brain activity need not be linear…
Sure, that's fair. But, do you think it's really just a low-D manifold curled up many times?
Why would you actually want a low-D manifold for most real-world tasks?
I could imagine lots of reasons they'd be seen. Robustness, predictability of future possible states, etc. Maybe you don't even want it, it's just a necessary consequence of wiring billions of neurons up and getting a somewhat-stable solution.
A simple bowl could be embedded in a 10^9 dimensional space. But naturally I don't think anyone argues "the brain is 10D" (or at they shouldn't).
One of the reasons we wrote our recent perspective (
www.nature.com/articles/s41...) was to try and argue that the talk around manifolds should be more about what it buys us conceptually as a framework, not about "is it low-D".
Well, do please note: my comment above is not intended as a slight against manifold-based approaches.
The point is merely that the idea that those manifolds are super low-D is not justified by the theory or data, at this point, I'd say...
Yes this I agree with 100%! I agree that the low dimensionalities we've seen to date are likely because we collect data in simple/limited behavioral regimes and short timescales.
Dec 12, 2025 19:34I 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.
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.
High-dimensional neuronal activity from low-dimensional latent...
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...
