Oddball First Results (also more RFs and ORI)

[maierav]

There was a Claude.Science hackathon last month that provided near unlimited tokens, so I took advantage of that. Here is the resulting repo:

https://github.com/maierav/ai_oscp_neuro

At the very least, the README.md might be useful to you to add to your prompt/context window (or as a skill file of sorts) to quickly get your own AI agents up to speed regarding the dataset.

Some interesting tidbits:

1. Starting from scratch, we ended up with receptive fields matching prior posts on this forum:

2. Same for the direction/orientation tuning:

So, while AI code obviously cannot be trusted, the fact that we get converging results using different agents is encouraging. This matches Sarah's idea of red/blue teaming the AI code: Task Codex and Claude Code with the same analysis and see if they produce the same result. If so, confidence in the code increases.

3. Pilot analyses of the various oddball types are both promising and interesting:

As expected, the oddball findings will require a more fine-grained look to decide on what is common and what is not across the oddball types, techniques, sessions, animals, areas, layers, neuron types etc.

So far, it looks like very clear results for the Neuropixels, which makes sense given the large sample size: positive effects (more spiking) for all types of oddballs.

There are also interesting and clear results for the Mesoscope, but they do not fully align with the Neuropixels in V1.

SLAP2 is less clear (no strong effect so far), but we only have 2 sessions that match the other techniques available for analysis so far.

I suggest to treat all these AI-produced results with a grain of salt for now. The probability that there are bugs in the code is rather high. Since I have been gifted the tokens, I put several agents on an audit and QA task for the repo, which is ongoing. The repo might change as bugs get detected. Eventually, I will manually inspect the code.

The analyzed sessions used what was available on DANDI as of yesterday:

[jeromelecoq]

RF and orientation tuning is very cool!

We will need to start thinking about sampling bias.
Meso sample more broadly and has way more inhibitory neurons notably.
great discussion to have next week!