• Date: 2026-08-18
• Time: 09:00AM (PT)
• Location: Teams Meeting

Agenda

Round-table update on progress around figures 6, 7 and 9 for the Predictive processing data release manuscript

• Carter Peene will present a brief update on our data collection tables.  
• Alex Maier will then lead with a first pass analysis of oddballs and tuning.  
• After that, we will be open to discuss any analysis progress made within the last few weeks. 

Meeting Recording

.

Meeting Notes

Data Collection Tables and QC Updates: Carter Peene provided an update on the data collection tables, including both interactive and static views, QC tags, and synchronization with the public spreadsheet, with Marcel Raabe and Sarah Ruediger Lee raising questions about session uploads and display features.

QC Tags and Session Status: Carter explained the use of QC tags to indicate reasons for session failures and clarified that some tags, such as Z drift, may appear in passing sessions. The criteria for session failure and QC tag application will be updated in the table captions to improve clarity.

Interactive and Static Views: The interactive view allows users to search and filter sessions, while the static view summarizes worksheet inputs. Carter noted that the data access table is synchronized with the public spreadsheet and provides compressed links to session assets, including DANDI links.

Session Upload Status and Discrepancies: Marcel inquired about the status of SLAP2 session uploads, noting that the most recent session on DANDI was from December and only contained one zebra block. Carter responded that three non-pilot SLAP2 sessions remain to be uploaded, with technical issues being addressed, and confirmed that the Google Sheets are periodically updated.

Suggestions for Table Improvements: Sarah suggested adding hover-over date information to session blocks in the display to facilitate cross-checking for users like Marcel. Jerome and Carter agreed that this feature could be implemented to make the tables more dynamic and user-friendly.

Future Data Collection Plans: Jerome mentioned plans to collect more data in the fall to increase session numbers, as agreed with the Kaspar team, and outlined ongoing efforts to upload and process voltage sessions for SLAP2.

AI-Assisted Analysis of Oddballs and Tuning: Alexander Maier presented a first-pass analysis of oddballs and tuning using AI tools, detailing the methodology, preliminary findings, and code review strategies, with input and questions from Farzaneh Najafi, Marcel Raabe, and Sarah Ruediger Lee regarding scientific interpretation and robustness.

Methodology and AI Tools: Alexander described the use of Cloud Science and other AI agents to analyze the dataset, emphasizing the role of sub-agents as reviewers to catch hallucinations and bugs. The analysis was conducted using multiple paradigms and recording modalities, with results and code made available in a public repository.

Preliminary Findings Across Paradigms: The analysis revealed that three out of four oddball paradigms (feature-oddball, sequence, duration/timing) elicited positive prediction-error responses, while sensorimotor oddball did not show a detectable effect. Cross-modality comparisons indicated consistent directionality but varying effect sizes, and subject-level robustness was assessed.

Scientific Interpretation and Caveats: Farzaneh raised concerns about interpreting omission responses as oddball effects, referencing her upcoming paper and emphasizing the need for careful manipulation and analysis before drawing conclusions. Alexander and Farzaneh discussed the importance of multiple analyses and convergence to ensure robustness and avoid premature claims.

Code Review and Reproducibility: Alexander explained his approach to code review, involving both AI and human checks, smoke tests, and convergence across independent analyses. The repository includes reproducible notebooks and an offline pytest suite to ensure correctness and provenance, with open-source code for community validation.

Laminar and Area Analyses: The group discussed laminar and area-specific analyses, noting challenges in layer assignment and differences between modalities. Jerome and Alexander agreed on the need for dedicated validation of depth and layer matching, cautioning against over-reliance on metadata and emphasizing biological variability.

Coordination for Data Release Paper Figures: Jerome Lecoq proposed a plan for participants to submit two slides ahead of the next meeting, focusing on figures relevant to the data release paper, with Sarah Ruediger Lee clarifying panel requirements and Alexander Maier asking about code versus graphics for figure contributions.

Slide Submission Plan: Jerome suggested that each participant submit two slides before the next meeting, highlighting material for the data release paper, to streamline discussion and ensure everyone has an opportunity to present their work.

Figure Requirements and Integration: Sarah asked whether individual panel labels should represent single or multiple plots, particularly for quality metrics in electrophysiology. Jerome clarified that multiple plots per panel are acceptable and encouraged submission of slides that aid manuscript integration.

Code Versus Graphics for Figures: Alexander inquired whether figures should be accompanied by code or just vector graphics. Jerome responded that ideally, figures should be generated by code that can be integrated into the manuscript and run from S3 or NWB files in DANDI.

Focus on Manuscript-Relevant Figures: Jerome emphasized the importance of figures for unit identification, oddball analysis, and receptive fields, requesting that participants prioritize material that supports the data release paper and enables follow-up studies.