Read on Twitter
Thread on recent papers about how mouse V1 populations code sensory information.
The paper I had looked at most closely was Kafashan et al. from Harvey and @drugowitsch. The data effects I focused on were (1) sublinear scaling of info, and ... https://www.biorxiv.org/content/10.1101/2020.01.10.902171v1
The paper I had looked at most closely was Kafashan et al. from Harvey and @drugowitsch. The data effects I focused on were (1) sublinear scaling of info, and ... https://www.biorxiv.org/content/10.1101/2020.01.10.902171v1
(2) f' direction (stim direction) not in the same direction as the principal noise direction. It seemed to me that all three papers (w/ Stringer et al. bioRxiv, Rumayantsev et al. Nature, came up with similar results for these. That's good!
@computingnature showed motor info was orthogonal to visual information, which to me is consistent with all the above. Rumayantsev found 3rd biggest component was noise in stim dir; ...
Kafashan found info-limiting noise small and distributed over many PCs (re: citations, I didn't look carefully at who was cited, but Carsen paper seems to deserve citation.)
I don't have a horse in this race, other than on behavior, having made est of mouse ori change-detection thresholds (decent: 6-10 deg for a small Gabor; Lindsey Glickfeld in a newer task w/ controlled adaptation found worse; but incr. size probably much improves thresh).
I felt good about all three papers because to me all the results are consistent with what Pitkow, Pouget and collaborators predicted for how stim-direction (info-limiting) correlations restrict sensory coding.
Q: it would apply to all 3 similarly, but there might be GCaMP thresholding effects- perhaps consequences can be simulated? B Averbeck recent finds qualitatively similar scaling (4 Utah electrode arrays), but not V1, so decoding perf not directly comp.: https://www.jneurosci.org/content/40/8/1668
And what's said on this assumes a feedforward model (i.e. noise comes from input fluctuations), and it could well be that there is feedback noise in the stim direction.
Last, it's important to know perf diff between a diagonal/independent decoder and one that accounts for correlations. Nice that there were some quantitative comparisons in the papers.
cites: Rumayantsev https://www.nature.com/articles/s41586-020-2130-2?proof=trueMay%252F
Stringer et al. bioRxiv https://www.biorxiv.org/content/10.1101/679324v1
In sum I felt, after looking at all this work, that we've made some progress understanding neuronal population coding.
Also, I screwed up first tweet - it's @jdrugowitsch
