So glad this technology is reaching more and more research labs! It's an incredible tool to crack neural codes!!

100% agreed! Thanks so much for sharing :)

Also, huge thanks to the Ho Yin Chau, @apalmigiano.bsky.social & @kenmiller.bsky.social for the many MANY discussions over this complex data and the possible circuit computations! It's been a huge learning opportunity 💙

Also big thanks to Dr. Lucia Rodriguez (x100), @danfeldman.bsky.social @amarinburgin.bsky.social and Dr. Kaeli Vandemark for their super valuable feedback on the manuscript.

Huge thanks to all the authors!!, especially @lamiaeadm.bsky.social who designed and built this powerful all-optical system, and to the great Adesnik Team!!

We think feature-specific recurrent inhibition may be a general cortical strategy to minimize redundancy, suppress ambiguity, and sharpen internal models of the world.
Read the full story: www.biorxiv.org/content/10.1...

Our results:
🔘Identify a feature-specific PC→SST→PC motif
🔘Show how it can switch from completion to cancellation (unifying previous findings)
🔘Demonstrate how feature-tuned recurrent inhibition sharpens cortical codes

So WHY is the brain wired with a like-to-like inhibitory loop?
Stimulating co-tuned SSTs while showing their preferred visual input:
-Reduces evidence for flanking orientations (consistent with explaining away)
-Preserves evidence for correct orientation
-Boosts discriminability

In fact, directly activating co-tuned SST ensembles alone is sufficient to remove input-matching representations in the absence of visual input.

Feature completion can be explained by the well-known like-to-like PC-PC connectivity in V1, but where does the feature-specific suppression come from?
-PCs recruit co-tuned SSTs (not PVs)
-SSTs, in turn, suppress co-tuned PCs → a “like-to-like-to-like” inhibitory loop

Same microcircuit, opposite computations, depending on input sparsity. This partially reconciles previous contradictory findings using similar tools. But how does this happen?

Using all-optical physiology in awake mice we photostimulated orientation-tuned PC ensembles in V1 in the absence of visual input, and we found:
Small PC ensembles → dominant feature suppression
Large PC ensembles → dominant feature completion

This means that both excitatory AND inhibitory connections in the cortex are highly structured: They store information (statistical regularities → “an internal model”) that can be used during sensory processing.

We show that this dual capacity is present in the same circuit with two components:
(1) Like-to-like connections between PCs (for pattern completion)
(2) A newly discovered circuit motif: Reciprocal like-to-like connections between PCs and SSTs (for pattern cancelation)

Brain circuits can use learned statistical regularities to enable completion or cancelation of predictable signals, but how?