Exploring Low-Dimensional Structure in SLAP2 Population Activity

[JJYma79]

Hi everyone,

Following Jerome's guidance and suggestion, I organised my current analyses into four figures and would greatly appreciate feedback from the community.

For Figure 1: Shared temporal bases
To address the concern that apparent population structure might reflect noise or ROI-specific fluctuations, we asked whether trial-averaged population activity could be captured by shared temporal bases.

For Figure 2: Shared population covariance
To address the concern that PCA might introduce artificial structure through its orthogonality constraints, we used Factor Analysis to test for a reproducible low-dimensional shared covariance structure in ROI activity, with PCA included as a comparison.

Next, Figure 3: Task-related latent population structure
We asked how task variables, including block context and stimulus orientation, influence the shared latent structure of population activity.

Then for Figure 4: Validation of condition-dependent population structure
We asked whether condition-dependent population structure remains evident when activity is represented as ROI–time patterns, and whether the resulting low-dimensional patterns are broadly distributed across the population.

The figures contain my current interpretations and main readouts.

I would greatly appreciate any feedback, comments, or critiques.

Thank you very much for your time and suggestions！

[jeromelecoq]

Hi Jessie,
I am going through your figures. This is a lot of work and there is a fair amount of material. If you were interested in presenting at our weekly meeting, you would be welcome to do so.

In the interim, I read that :
1/ You have validated that PCA components are robust and FA yields the same representations.
2/ You found 3 main modes of activity upon trial presentation : a strong positive response, a strong negative response and a biphasic response
3/ In your analysis, there does not seem to be difference between DMD1 and DMD2 representations?
4/ You found that the orientation blocks yield the largest amount of variance accounted for by the lower dimensional model.
5/ 90Deg stimuli trials is over-represented after PCA/FA. is that right?

Remind me: Did you run PCA across all blocks in one go?