Discussion of Receptive Fields Analysis and Alignment On Figures for Data Release Paper

[Dedalus9]

As mentioned during our last meeting, I think it would be good if we got aligned on our respective receptive field analysis figures for the data-release paper, specifically as it pertains to figures 7 and 8 of that paper.

Current Receptive Field Analyses

From what I recall, receptive field analyses have been conducted by @RaabeM, @lrudelt, and @maierav.

• The receptive field analysis of @RaabeM / @lrudelt was focused on the zebra stimulus. Their analysis focused on single ROIs (data and receptive Fields figure for Data release paper (need feedback) #151 (comment)).
• My receptive field analysis looked specifically at the receptive field mapping block (Discussing the Ophys QC Pipeline #156 (comment)). It focused on population-level stimulus responses in the mesoscopic data. I also looked at representational stability across recording sessions and subjects for each layer.
• @maierav's analysis (Oddball First Results (also more RFs and ORI) #168 (comment); Some Analysis Progress #145 (comment)) is focused on receptive fields during the oddball task.

Alignment Across Modalities

If we would like to repeat our respective analyses across modalities (i.e., NeuroPixels, Mesoscopic, and Slap2), I think it would be useful to align on how our analyses overlap and complement one another. If we're making figures for the data release paper, our approaches should be complementary rather than redundant.

Overall, I think the emphasis should be on validating the stability of receptive fields across recording sessions and subjects. I'm curious to hear everyone's thoughts on this.

Meeting

Perhaps we can all meet next Monday? @carolinehaoud would also like to join this meeting. I could also meet on Friday if that works better.

Missing Analyses

I may have missed some receptive field analysis work. If so, could you include a link to your analysis in the comments? I may not have found every receptive field analysis in the discussion threads.

[RaabeM]

Hey all,
Thank you for the thread!

I think for each modality we should pick one session and show the mean RF and a view specific RF and tuning examples. Multiple people already worked on this with different approaches. So, we should compare results and methods and see whether we can converge to one solution, which we want to write up in the end. Moreover, the repeatability/stability of neural responses are interesting and should be shown additionally. With your metric, @Dedalus9, this can also be done for multiple/all sessions.

Sessions to analyse

Just a suggestion on which sessions we could test:

• Ephys: sub-830794/sub-830794_ses-ecephys-830794-2026-01-26-12-02-05_ecephys - motor cohort v2
• Meso: sub-832700/sub-832700_ses-multiplane-ophys-832700-2026-01-24-12-06-12_ophys - motor cohort v2
• SLAP2: sub-829704/sub-829704_ses-829704-2025-12-18-10-57-36_image+ophys - motor cohort

During these sessions no errors/problems occurred and I know that some of us already analysed some of these sessions. But please note, that all the sessions are from mice from the motor cohort. Do you see any problems with it?

Meeting

Unfortunately, I will not be available next Monday. Let's do a poll for scheduling. I used Mo. and Fri. as possible weekdays. Are there more days I can include? I am not aware of all our timezones , therefore the list of times is rather long. Please note, that all times are in CEST.
➡️ Here the link to the poll.

Current state

I think you catched the relevant threads. I just wanted to add, that I also analysed the RF mapping block. You can find a comparison to the zebra analysis in this presentation.

[Dedalus9]

I agree that for each modality we should pick a session and show the receptive fields of individual ROIs. I also agree that an analysis of the stability of representations across laminae and between subjects would be relevant for a data-release paper, particularly since they tell the interested researcher whether it's okay to analyze the data (e.g., representations) when averaged over trials or whether they should also look into trial-specific variability. I could also repeat the same representation stability analysis with respect to the zebra stimuli using your approach.

However, I do also think that a population-level analysis would be complementary to all of this. For instance, a cortical area may output to another downstream cortical area the integrated activity of population of neurons rather than that of individual ROIs. This would make looking at patterns of population-level activity relevant for an analysis of predicting processing.

If we're in agreement on this, the receptive field analysis would include four types of figures for zebra/gabor stimuli across all modalities:

• Figures of individual ROIs
• Population-level response patterns
• Stability of representations across subjects
• Stability of representations across recording sessions for each subject

Also, I filled out the poll!

[maierav]

Looks like all of us might have time this Friday. Shall we commit?

[RaabeM]

Thank you all for voting!

The meeting will take place on Friday 21'st at 3 pm (CEST).

Let's meet here on Zoom ➡️ https://uni-goettingen.zoom-x.de/j/64723395080.

@maierav @Dedalus9 @Sarruedi @lrudelt @carolinehaoud

[Dedalus9]

#156 (comment)

[maierav]

Here is an attempt of adding unique IDs to the receptive fields, so we can compare them using our different techniques:

ID Key used below:

Six best-fit RFs per technique (rows: Neuropixels spikes, mesoscope somatic ΔF/F, SLAP2 dendritic glutamate), each panel titled with unique ID:

Code: https://github.com/maierav/ai_oscp_neuro/blob/main/notebooks/rf_labeled_examples.ipynb

[Dedalus9]

Should we plan on submitting a pull request this Friday? That is, to submit our work for the data release paper

[RaabeM]

Let us discuss that on Friday. I think we should commit when we converge agree on a final version. However, I think you could commit your similarity measure also earlier.

[koweiss]

I might have missed that but it would be great if @jeromelecoq provide some definition of an inclusion/exclusion criteria for analysis/figure of receptive fields characterization in this data release paper.

[RaabeM]

We want to select the units/ROIs with the most significant RFs. Any noisy unit/ROI should not survive the statistical testing.

[jeromelecoq]

We were planning to have a cross-modality RF figure so this seems like it checks out for inclusion yes.

[RaabeM]

Dear all,
I am posting the top 10 results for the three sessions we agreed on. You can find all plots and the regarding csv files for comparison here.
The plots look like the one I am posting below. I am showing top 10 RFs for each modality sorted by the most significant zebra/gabor results. I still have to do some last checks. For example, I have to ensure, I did no mistake with the orientation of the gabor results. The SLAP2 zebra analysis is still running. I hope I can show its results tomorrow. However, the good news for SLAP2 is, that I find nice RFs in both channels.

[RaabeM]

Edit: Just an update on the corrected results presented during our last meeting. RFs match between zebra and gabor stimulus. Also confirmed by Alex.

[Dedalus9]

Following @maierav's advice, I looked only at the x-/y- tuning of ROIs since we expect for orientation tuning to be more variable over small distances in the visual cortical areas. Toward this, I averaged the response of each ROI to x- and y- stimulus presentations. To make sure that my approach is agnostic to what ROIs are imaged on any given recording session, I looked at how similar the covariance matrices were across recording sessions for each imaging depth. These covariance matrices were computed with respect to stimuli rather than ROI. In other words, these covariance matrices were stimulus x stimulus rather than roi x roi. Central Kernel Alignment (CKA) is the method I used for computing this similarity.

In the figure, the top row corresponds to mesoscope data. The bottom row corresponds to neuropixels data. The left column corresponds to VISp and the right to VISl.

What this shows is the stability of representations across imaging depths shows a consistent monotonic increase with imaging depth. This figure, then, accomplishes three things. First, it tells us how stable representations are for each layer with respect to mesoscope and neuropixels data. Second, it validates the similarity metric: we're finding the same monotonic increase in representational stability over imaging depths regardless of the imaging modality. Third, it tells researcher interested in representations whether it's okay to look at them over multiple recording sessions or to consider each recording session individually.

[lrudelt]

This is very interesting!

Great to see that the mesoscope recordings are more consistent across sessions, because these aim at targeting the same population.

Maybe one interesting reference would be splitting the trials of the RF mapping block in two halves (consecutive or randomly selecting trials) and to compute the CKA across these two blocks. Ideally these are highly consistent, at least for units that pass a certain quality threshold (e.g., p-value of their RF mapping response lower than some value). If Mesoscope achieves almost this value, then RF tuning is as similar between two sessions as it is between two halves of the tuning block within the same session, making a strong case that this is indeed the same neural population!

Beyond that, it would also give a good measure of how reliable the visual tuning is PER SESSION across recording techniques and neural structures (areas, layers, dendritic compartments), because visual tuning might be very different across sessions for ephys and SLAP2 recordings (these record different neural populations in the end).

[RaabeM]

There will be no meeting next week (04.09.), but the week after on 11.09.

[maierav]

Everything replicated! I get the same RFs using my vibe code (though these are averaged across all orientations):

Code is on the repo that I posted above.

[RaabeM]

Great! Very cool to see. :)

[maierav]

It's a bit scary. This suggests that AI hallucinations are not much of a concern anymore when it comes to agentic coding - even for more complex projects.

We are reaching the point where we can just trust the code (and hence the analyses and figures). And thus stop coding ourselves.

[RaabeM]

Well, I would not call the RF computation complex. But yeah, AI definitely became way more reliant. Comparing results on other tasks that need more interpretation of intermediate results would be very interesting! I find the discussion on how to deal with it very exciting. Currently, I still wouldn't trust it blindly.

[lrudelt]

Great to see the agreement!

For the paper figure, I was wondering whether it would make sense to show example RFs only for units from V1, potentially even all from L2/3, since this is comparable across recording methods (SLAP2 has only V1 L2/3 pyramidal cells). The very high correlations with Zebra >0.5 one most likely only finds in thalamus ;)