Agenda

Jerome will lead a discussion to consolidate our data release manuscript.

Meeting Recording

.

Meeting Notes

Receptive Field Analysis And Manuscript Figures: Marcel compared Zebra-noise and Gabor-patch receptive-field estimates across electrophysiology, mesoscope, and SLAP2 data, while Jerome, Stefan, Lucas, and Sarah discussed significance testing, visualization, data coverage, and figures for the manuscript.

Analysis Pipeline: Marcel explained that the Zebra stimulus is convolved with a library of Gabor wavelets varying in orientation, position, size, and spatial frequency. The analysis extracts the activation time course for each filter and position, correlates it with neural responses, selects the best-fitting filter, and moves that filter across the stimulus to estimate the receptive field. The pipeline also optimizes response-window duration and delay relative to stimulus presentation, and can be applied across trials and stimulus phases.

Cross-Modal Comparisons: For electrophysiology, Marcel reported that the top Zebra- and Gabor-derived receptive fields generally had aligned positions. Similar alignment was observed in mesoscope data and the SLAP2 green channel, although mesoscope correlations were lower than electrophysiology correlations. The SLAP2 red channel produced weaker results, with maximum correlations around the borderline of possible significance; Jerome cautioned that the red channel is recorded late in the session, when calcium-indicator bleaching and weak responses can reduce reliability.

Combined Receptive Fields: Stefan asked whether Marcel would combine significant filters and orientations into a conventional receptive-field representation rather than retaining only the best filter. Marcel confirmed that the current results use only the best-fitting filter and agreed that a linear combination of wavelets could provide a more classical receptive-field estimate. Stefan suggested evaluating either all relevant filters or filters passing a significance criterion, while noting that combining filters may be appropriate even before significance filtering.

Statistical Testing: Marcel planned to test Zebra receptive-field correlations using shuffled responses to estimate a null distribution. He initially proposed Bonferroni correction across units within a session, and Stefan recommended the less conservative Benjamini–Hochberg false-discovery-rate procedure instead; Marcel agreed to use that approach.

Manuscript Visualization: Marcel proposed plotting receptive-field centroid heat maps rather than only averaging complete receptive-field maps, because averaging can cancel spatially distributed responses. Jerome clarified that the method sums or bins the detected receptive-field centers, and suggested that a scatter plot of all centers could be an equivalent presentation. Sarah recommended a neutral gray baseline rather than black in the colormap so responsive regions stand out more clearly.

Data Coverage And Figure Planning: Marcel’s displayed examples used one session per modality, and Jerome requested a pull request containing the analysis code and recommended figures so the results can be integrated with other plots. Lucas asked about showing receptive-field positions and sizes across structures. Jerome said these comparisons require broader coverage than the current subset, and Marcel agreed to extend the analysis across sessions, preferably using event traces. Marcel also noted missing cortical-location metadata in some electrophysiology files; Jerome said the team would investigate and resolve the metadata issue.

Calcium Soma Annotation: Lucas asked how to identify soma ROIs in the SLAP2 red-channel data. Jerome explained that soma ROIs are generally larger and visually distinct from dendritic ROIs in projection images, and said the soma should be explicitly annotated in the manuscript. Jerome planned to discuss the annotation approach with Kaspar.

Cross-Session Representational Stability: Nicholas presented a CKA-based analysis of stimulus-wise receptive-field covariance across sessions and depths, prompting Jerome, Sarah, Lucas, Alexander, and Stefan to request controls for sampling bias, ROI selection, cortical context, and the interpretation of low similarity values.

CKA Method: Nicholas computed a stimulus-wise covariance matrix for each recording session using ROI responses to X/Y stimulus positions during the receptive-field mapping block. He then used centered kernel alignment to compare these matrices across sessions, with higher CKA values interpreted as greater similarity of the population-level stimulus representation.

Depth-Dependent Finding: Nicholas reported that receptive-field representation stability increased with imaging depth in both mesoscope and Neuropixels analyses. The proposed implication for dataset users was that representational drift should be considered, particularly for superficial recordings. Lucas noted that mesoscope CKA values were generally higher than Neuropixels values, potentially because mesoscope recordings aim to follow a similar population across sessions whereas acute Neuropixels insertions sample different units.

Context And Biological Controls: Jerome questioned whether the apparent superficial-layer instability reflected representational drift or the experiment’s changing behavioral contexts between sessions, which could alter inputs to superficial visual-cortex layers. Alexander additionally asked whether superficial recordings might be more affected by physical instability or movement. The group agreed that the surprising depth effect requires controls before being treated as a central result.

ROI Selection: Sarah asked whether Nicholas had filtered for units with clear receptive fields and whether nonselective units could confound the CKA result. Nicholas said the current metric was population-level and not restricted to units with individually clear receptive fields, but agreed to repeat the analysis after filtering units using Marcel’s receptive-field quality criteria. Sarah said this would make it easier to determine which response properties drive the depth-related change.

Sampling Bias Concern: Stefan cautioned that CKA assumes broad and unbiased sampling. If recordings sample different retinotopic locations or restricted subsets of tuned units across sessions, CKA could fall substantially even when the underlying representation is unchanged. Stefan recommended testing this with controlled simulations, such as passing the dataset’s stimulus types through early convolutional layers of a network and shifting the sampled spatial region or receptive field, then examining whether CKA decreases artificially.

Follow-Up Analysis: Jerome summarized that Nicholas should incorporate the statistical and methodological feedback, perform additional controls, and present the analysis again. Nicholas agreed to investigate the validity of CKA under biased sampling and to extend the work toward individual-ROI analyses where possible. Individual-unit tracking is feasible for mesoscope data but not straightforward for Neuropixels data.

Mesoscope Registration: Nicholas asked whether metadata already identified the same ROIs across mesoscope sessions. Jerome clarified that the sessions were acquired in the same approximate location but had not yet been registered to one another. The team has supporting registration code that is not yet packaged, and Nicholas could perform the registration and contribute the resulting code for broader use.

Analysis Code Validation And Deferred Updates: Alexander reported that the AI-based and manual receptive-field implementations produced matching results, while the planned stimulus-statistics discussion with David was deferred to the following week because of time.

Code Comparison: Alexander said his comparison of Marcel’s implementation with the AI-based code found essentially the same receptive fields. Based on that agreement, the team could proceed with the AI code, which Alexander said was further advanced in some parts.

Deferred Stimulus Discussion: David had prepared material on stimulus timing, mismatch frequency, and running statistics in the sensorimotor context block. Jerome and David agreed to postpone that discussion until the next meeting.

Upcoming Data Release Activities: Jerome announced an October 8 SFN poster opportunity and an early-October multiplex-FISH presentation that will add genetic cell-type information for a subset of mesoscope sessions to the planned data release.

Conference Poster: Jerome said the SFN poster was submitted for October 8 with as many authors as permitted and encouraged participants attending the conference to use the opportunity to meet and discuss the work.

Multiplex-FISH Data: Jerome announced that the team had extracted cell-type information from a subset of mesoscope sessions using multiplex FISH. The resulting genetic information for imaged cells is intended to become part of the data release, and the team planned to present the dataset in early October before packaging it for distribution.

Next Meeting: Jerome said the next meeting would return to David’s stimulus-related material and any other submitted presentations, noting that participants could continue to add topics to the discussion forum.

💬 Join the discussion for this page on GitHub