Discussion: meetings/2025-09-09

[jeromelecoq]

Where to discuss the presentation from Sophie and Maxwell.

[jeromelecoq]

I thought it would be helpful to explain in more depth why I think this stimulus could be valuable to our work. I think there were a few questions from Sophie today. Perhaps I should have better explained this at the onset.

Our goal with this predictive processing study is to measure how neurons are making predictions across different contexts. In particular, we postulate that when sensory context is changed, the mechanisms leveraged by individual neurons to make predictions of future stimuli fundamentally changes. Our review of the literature has highlighted disagreement across studies in how inhibitory neurons or different neuronal pathways are recruited (among others mechanisms).

Fundamentally, we think it is due to difference in stimuli design. It is unfortunate but predictive processing has been tested with so many different tweaks in the stimuli that it is difficult to reconcile those difference to build proper models. We think the cortex has many tools in its toolkit to make predictions: Sensory-motor stimuli are leveraging input from motor cortex coming in superficial layers while quick repeats of a grating can trigger prediction solely based on local recurrent information, perhaps even at the level of single cells, within dendrites.

We aim to help consolidate the difference in the literature by collecting a dataset that span 4 different sensory context in the same mouse. and 3 recording modalities (2p, Neuropixels and Slap2) across different mice.

Our design leverages drifting grating heavily so it is important to this work that we have good baseline system identification (as Kaspar mentioned) across those stimuli. We see the visual cortex as leveraging those baseline system responses to make predictions. So much is known about neurons in the visual system in the mice: It is important we know which neurons have the same receptive fields.

Receptive fields also are a key QC metrics to know our recordings are working well, so we typically have a 10min Gabor receptive field block added to all of our sessions. It would be sad to do so much work to later realize that the neurons we are recording are not receiving inputs from our stimulus screen.

This is where our interest in your stimuli came about : Quick baseline system identification. In the rest of our sessions, we present repeated stimuli that pushes the system to adapt and learn using those baseline responses.

[jeromelecoq]

Does the rig geometry come into play in your generation? All of our stimuli are warped to flatten spatial frequencies. So I assume we need to adjust based on the distance to the eye and band of spatial frequencies presented ?

[mwshinn]

We aren't currently adjusting for this, but the zebra noise video could be adjusted after it is generated. Since the analysis is performed on the raw frames of the video, perhaps this would mean adjusting the video that is displayed but not the one that is analysed with the neural data. (@skriabineSop thoughts?)

[jeromelecoq]

1/ our warping is applied by bonsai after so it does not need to be in video generation
2/ How does the distance to screen influence the zebra movie you play?

[mwshinn]

Current parameters were tuned for rigs that have a 10 cm distance to the screen.

[skriabineSop]

So basically the only thing that the distance to screen distance would influence is the number of pixel on screen corresponding to a visual degree. And how many degrees of the visual field your screen covers. If you have a picture of your setup to share with us it would maybe help giving you advice on which parameters to use?

[Sarruedi]

I think there may have been some confusion about the potential reason for using the zebra noise stimulus. My view is that it could provide a very efficient way to parameterize the large stimulus space quickly. Our actual paradigm would then still make use of a much narrower stimulus set, based on drifting gratings at particular orientations as planned.

Personally, I like the idea of the zebra stimulus as a way to access neurons that may have different tuning preferences than those we typically capture with the standard RF mapping protocols. This would allow us to test whether the properties of cells we typically would not even register also change as a function of the experience we subject the animals to. It could be especially useful to use the zebra stimulus for a broad characterization at the very beginning of an experiment, and potentially again at the end of the session, as we have done in the SLAP2 design.

[jeromelecoq]

That being said, It is possible that the responses are so reliable "because" the contrast is so high.

[jeromelecoq]

But that comes at the cost of NO activation of some inhibitory sub-populations.

[skriabineSop]

there is actually a stimuli called Trippy that is used in there: https://www.nature.com/articles/s41586-025-08790-w. It s not very documented, I think it s made of cosine of gaussian noise. We tested it against the zebra noise, and it appears to drive neurons less reliably than zebra. We think, as you said that it s probably due to the absence of sharp edges.
Something that we were planning is to have an option to add some gaussian blur to the zebra noise to get smoother transition between the stripes, while keeping its orientation invariance and the spatial decorrelation qualities. Might be what you are looking for ?

[maierav]

Here's a bit more on Trippy (though no code):
https://github.com/cajal/trippy-monet?tab=readme-ov-file

If there is interest, I can try to apply a filter to Zebra to turn it into something like Trippy.

I agree with all points raised. There are both pros and cons with either choice.

[jeromelecoq]

The more I think about it. The more I am convinced we need to present variants with high and low contrasts of one of those. The trippy stimuli in the connectomics is interesting because it creates a bridge with highly valuable EM data.

[severine2305]

Hi @skriabineSop, great talk yesterday! Tell me if I missed it but did you ever compared directly the tuning properties of each cell obtained with the zebra stimulus vs the properties you obtain with drifting gratings and sparse noise? thank you.

[skriabineSop]

Hi Severine ! Thanks ! yes I did some comparison. to summarize briefly, the retinotopic maps were similar to what you would get with sparse noise, except that where you would need 20 or more minute of sparse noise to obtain them, you would only need 5 min of zebra noise. Same for drifting grating comparison. if I remember correctly, about 90% of neurons had the same tuning when measured with both standard and zebra noise stimuli. We believe the remaining ones are just not properly driven by sparse noise or gratings because they may have a more complex receptive field . If you want to see the data you can check out Supplementary figure 6 of the paper [https://www.biorxiv.org/content/10.1101/2025.07.19.665666v2.full.pdf]

[jeromelecoq]

Yes, I think you are getting part of my point. I also meant which I would expect the current design of zebra to recruit subsets of cell. So I was curious whether you have tested some of those ideas. The Gaussian blur could be useful but I think lowering the contrast might be cleaner *Jérôme*

…

On Thu, Sep 11, 2025 at 4:21 AM skriabineSop ***@***.***> wrote: there is actually a stimuli called Trippy that is used in there: https://www.nature.com/articles/s41586-025-08790-w. It s not very documented, I think it s made of cosine of gaussian noise. We tested it against the zebra noise, and it appears to drive neurons less reliably than zebra. We think, as you said that it s probably due to the absence of sharp edges. Something that we were planning is to have an option to add some gaussian blur to the zebra noise to get smoother transition between the stripes, while keeping its orientation invariance and the spatial decorrelation qualities. Might be what you are looking for ? — Reply to this email directly, view it on GitHub <#111 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATAHT4KZQRUOG6BONS7HST3SFLLPAVCNFSM6AAAAACGBO26KGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMZXGMYTSNI> . You are receiving this because you authored the thread.Message ID: <AllenNeuralDynamics/openscope-community-predictive-processing/repo-discussions/111/comments/14373195 @github.com>

[skriabineSop]

You were saying that some inhibitory subcircuit might be recruited with lower contrast, if I remember correctly ? Since in the 2p you can only access excitatory cells, we did not really look into the inhibition question. But it s true it might become central with ephys.