Adil Khan @adilgkhan@qoto.org

V1 responses are enhanced with visual attention of various forms. An appealing hypothesis about how this happens is: Top-down inputs activate VIP interneurons, which disinhibits pyramidal cells through SOM cell inhibition & causes enhanced/modified responses to the same stimuli.

Ideally to test this, you would need to make an animal do an attention task, find attentionally modulated cells, and transiently switch off or switch on local VIP cells while simultaneously looking at responses of the attentionally modulated pyramidal cells.
And since any circuit manipulation has complicated effects on different cell classes, ideally you should know how multiple different cell classes respond to attention AND to the VIP manipulation in the same tissue (e.g. pyramidal, SOM and PV cells, in addition to the VIPs). Thats exactly what Dylan and team did!

First, we established that attention leads to strong modulation of stimulus response selectivity, and separately, VIP activation leads to strong modulation of pyramidal cell activity. But when we put the two together, there was no interaction. ie, attention and VIP led to independent modulations.

This result was confirmed by silencing VIPs, which left the attentional modulation untouched. So attention signals do not pass through VIP cells. In fact, we found that at the population level, the two modulations were orthogonal.

And because we had measured the activity of VIP, SOM, PV and pyramidals simultaneously, we could show that the changes in different cell classes and their interactions was distinct between attention and VIP modulation.

But how is it possible to have a strong modulation by attention AND by VIP on the SAME cells, but the two not interact? With our amazing collaborators Katharina Wilmes and Claudia Clopath, we found a very specific circuit architecture that can support this property.

Please read and share the preprint, which has much more in there. This was a great team effort led by Dylan Myers-Joseph, and with tremendous contributions from @k47h4, Marian Fernandez-Otero and Claudia Clopath

(11 of 11) This work was a great team effort, driven by Nick Cole
with tremendous contributions from Matt Harvey, Dylan Myers-Joseph and modelling of the behaviour by Aditya Gilra. Feedback welcome! Please RT

(10 of 11) If you've made it this far, check out the paper which has some other cool findings, particularly context-switching RL models. Also, I have found that its easy to misunderstand this task, so please reach out if any doubts.

(9 of 11) VIP activation almost abolished the mismatch signal! (Since the manipulation was only in the imaging site, not bilateral, this didn't affect behaviour). So VIP cells are key to generating the prediction error signal. This effect (riding on disinhibition by VIPs) constrains circuit models of mismatch
Image

(8 of 11) Anyone who's trained a mouse will know, if they don't get it right a few times they will do exploratory licks. Mice with ACC silenced ignore tens of trials of visual stimuli, rewarded if they would lick, but they are stuck in the odour block rule: ignore visual, wait for odour...

And once they do eventually switch, they do near-perfect performance the rest of the block, with ACC still silenced! We also did a fun single-trial Un-silencing experiment to nail this point. So ACC is truly a gate, allowing transitions between cognitive states (and PL is not)

Finally, what is the circuit for comparing prediction and reality? Amazing prior work (Keller lab) shows VIP & SOM interneurons are key. So we activated VIPs while measuring mismatch signals. All-optical method so we are sure the manipulation caused disinhibition as expected

(7 of 11) Coming back to the ACC, are these mismatch signals actually needed for switching behaviour? We silenced ACC: the mice showed a major deficit in switching, but perfect accuracy once switched, each time! Neighboring prelimbic didn't have any such effect. This is something to behold ...

(6 of 11) If the mismatch signal is so important to behaviour, is it broadcast across most of cortex? we used widefield calcium imaging to map it across the dorsal cortex, and found it to be beautifully restricted to frontal areas (above ACC/PFC). Similar to primates :)

(5 of 11) Crucially, the amplitude of this signal is larger when the NEXT trial is going to be a correct one-shot transition (compared to transitions where it makes 2 or 3 errors before switching). The mismatch amplitude matters to upcoming behaviour. An RL model showed cool similarities

(4 of 11) ~10% of all ACC neurons show this, a big number considering PFC responses. And we didn't see this signal by chance/movement artefacts, eg.identical recordings from V1 didn't show it. Also this signal is different based on where cells project to: mismatch signals are excluded from ACC neurons projecting to the striatum

(3 of 11) This is not just a surprise, like oddball tasks. This is a prediction mismatch which has demonstrable behavioural consequences - the mice switch behaviour using it. A complete switch in mental rules happens at a well defined moment - the mouse proves this to us by its behaviour

We recorded from the ACC, and aligned activity to the moment the odour was expected to arrive but didn't, and found neurons with a clear response - to the NON occurrence of an expected odour. Remember nothing happened at this moment - only the expectation violation.

(2 of 11) When they expected an odour (rule2) which didn't arrive, they switched rules, and started responding based on rule 1, the very next trial, and stayed very accurate for the rest of the block, and did this over and over allowing us to zoom in on the rare moment of prediction mismatch!