Yeah, I remember hearing about this by @suryaganguli.bsky.social a few years before their preprint, I think from @sasolla.bsky.social

And I think the picture we have now that (at least for embedding) dimensionality = f ( task "complexity", brain area )

Yes I think there are many more meaningful --pun intended-- questions that we're making progress on than that of dimensionality

Indeed; we're a young field!

Larger for some regions like striatum when compared to M1

(Also the estimated linear dimensionality tends to increase with the number of neurons, which is more relevant now that we have neuropixels...)

Yes it's fascinating that a plane works so well, but a curved surface still works better (we showed that one the preprint below). This difference becomes stronger for more complicated tasks as one would expect but the most interesting bit is that nonlinearity is ...

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

this looks very interesting, adding it to my to-read list!

also, I think that was with PCA —correct me if I'm wrong—, and PCA estimates the embedding dimensionality of a manifold not its intrinsic dimensionality, so it's an upper bound to the actual dimensionality

This paper explains it very nicely doi.org/10.1016/j.co...

this is a very interesting paper, but quoting from the paper:

"Analysis used a relatively high (30) dimensional space to ensure trajectory differences were not lost."

I'd argue that 30 is not very high compared to the millions of neurons in motor cortex...

what I mean is, is a naturalistic behaviour for V1, all images an animal may see in its lifetime? and for M1 all movements it may perform with a limb? Or are there principled subcategories of each of these?
I think this is quite complicated and would influence the answer to this question…

2) I think low-D should just mean —or at least it does to me— substantially smaller than the number of neurons, not "fewer than 10".

Also, I feel we —at least I— should have a better idea of what's a "behavior" because this is (obviously) required to measure dimensionality during behaviour

I think claims about low-D have received a bit too much attention, but I wouldn't throw the baby with the bath water:
1) very few studies estimate intrinsic connectivity, we mostly count embedding linear dimensions (PCs). These can be extremely different (e.g. a very twisted ring in high-D space)

Congratulations Dr Kashefi 🎉

Indeed
.. and sad to see Spain sinking after the positive trend

I hope you'll be recording this @dlbarack.bsky.social ?

actually it was just journal clubbed at ours

I wish I could!