A cognitive representation in primary visual cortex modulated by vision

TL;DR: Record from over a thousand neurons in rat primary visual cortex while the animal runs a memorized maze in total darkness, and you find that most of them are not silent — they fire in reliable sequences that track progress through the task, generalizing across physically different paths that share the same turn structure. Turn the lights on, and vision doesn't overwrite this signal; it multiplies it, like a gain knob. The paper's claim is that the classic driver/modulator story of sensory cortex — retina drives, brain state modulates — is inverted for a large fraction of V1: an internal model drives, and vision modulates. This is a preprint (bioRxiv, not yet peer reviewed), but the experimental design is unusually clean and the data and code are fully public.

Why this is a live question

The textbook V1 neuron is a feature detector: it has a retinotopic receptive field, prefers an orientation, and is driven by photons. Over the last fifteen years that picture has been eroded at the edges — locomotion, head movements, arousal, sound, and even estimated spatial position all modulate V1 responses. But nearly all of that work, especially the influential virtual-reality navigation studies (Saleem et al., Fiser et al.), studied how internal variables reshape visually evoked activity. The stimulus was always there; the internal signal was the seasoning.

The obvious missing experiment: take vision away entirely, in a real environment with real self-motion, and ask what V1 represents on its own. If the answer is "nothing much," the classic picture stands. If V1 turns out to carry a structured, task-defined representation with no visual input at all, then the question becomes what vision does to that representation when it returns — and that's exactly the manipulation this paper performs.

The setup: a maze you can run blind

Four rats ran a spatial-alternation task on a W-shaped track: start at the center arm, alternate visits to the left and right outer arms, return to center for reward. Correct behavior consists of four path types — center→left, left→center, center→right, right→center. The crucial geometric trick: these four paths pair up by turn structure. Center→left and right→center are physically different routes through space, but they share the same sequence of turns; likewise center→right and left→center. This lets the authors dissociate "where am I in the room" (allocentric place) from "how far into this leg of the task am I" (path progression) from "which turn program am I executing."

Each recording day included light epochs — monitors flanking the arms displayed gratings or dot patterns, with the stimuli on the two outer arms swapped between epochs (this swap becomes the paper's decisive experiment) — and a dark epoch with all visible light removed. The darkness control is done carefully: behavior was tracked under 940 nm IR at ~0.01 mW/cm², 48-fold below irradiance levels at which dark-adapted rats perform at chance on high-contrast visual discrimination. Four 128-channel flexible polyimide probes per animal yielded >1,200 V1 and >400 CA1 neurons, recorded simultaneously.

What V1 does in the dark

About 27% of V1 neurons behaved exactly as the textbook predicts: tuned to specific stimuli in light, essentially silent (<0.5 Hz) in darkness. Feature detectors, as advertised.

The other 73% stayed active in the dark, and a large fraction fired in strikingly reliable spatial patterns across repeated traversals. The key observation is what those patterns were locked to. Not a fixed location in the room: a neuron active midway through center→left was also active at the analogous progression point of right→center — a different place, reached by a different route, at a different distance from the start, but the same fraction of the way through a same-turn path. At the population level, neurons tile the full progression axis, and the sequence ordering learned on one path transfers to its same-turn partner (the off-diagonal diagonals in their Fig. 1D).

The obvious deflationary hypothesis is that this is all just motor signals in disguise — V1 is famously modulated by locomotion and head movements, and same-turn paths involve similar movements. The authors attack this with nested Poisson GLMs on held-out data: a motor model (speed, acceleration, head direction, signed angular velocity, angular speed, angular acceleration) versus motor-plus-progression. Adding path-invariant progression improved held-out log-likelihood for 99% of dark-active neurons (median ΔLL = 0.08 bits/spike). They also pit progression against two competing spatial codes: an allocentric place model (position on the maze, direction-free) and a distance-traveled model (linear distance from the start, shared across all paths). Path-invariant progression wins both (93% and 94% of neurons, median ΔLL 0.05 and 0.03 bits/spike).

The most direct evidence is decoding transfer. Train a Bayesian decoder on one path, test on another:

That ordering is the aha: the code transfers best to the path that is physically most different but structurally identical, and worst to a path that shares physical territory but differs in turn direction. Whatever V1 is encoding in darkness, it is organized by task structure, not by retinotopy or place.

Light as a volume knob: the stimulus-swap test

So what happens when vision returns? Single neurons mostly kept firing at the same locations within a path, but with amplitudes scaled up or down — differently in different maze segments (which contain different stimuli), and differently on the two same-turn paths (which now present different visual sequences). Quantitatively: circularly shifting the light tuning curve against the dark one shows overlap peaking at zero shift, with dark–light overlap reaching a median of 0.47 of the noise-ceiling-normalized explainable range. Meanwhile the path-invariance index drops in light, cross-path decoding transfer gets worse, path-specific place decoding gets better, and population dimensionality rises — vision injects path-specific degrees of freedom into a low-dimensional shared scaffold.

All of this is consistent with multiplicative gain modulation, but under a fixed stimulus layout, stimulus identity is confounded with location, so gain modulation, additive visual responses, and outright replacement all make similar predictions. The stimulus swap breaks the tie. Fit models on epoch AB, predict held-out epoch BA where the outer-arm stimuli have been exchanged:

The multiplicative model — a single tuning curve shared across dark and light, with one scalar gain per maze segment — beat the independent model (separately fit dark and light fields) for 74% of neurons and the additive model for 90% of neurons (median ΔLL 0.07 and 0.90 bits/spike), despite having the fewest parameters:

If you think in ML terms, this is FiLM-style conditioning: the internal task-progress representation is the feature map, and vision supplies per-context multiplicative gains. The authors' framing — a "task-state scaffold on which visual inputs act" — is essentially a world-model latent that sensory evidence sharpens rather than generates. In darkness, two same-turn paths collapse onto one low-dimensional trajectory; stimulus-specific gains are what pull them apart and disambiguate otherwise-equivalent states.

The hippocampus connection

Where does an internally generated, task-defined code in V1 come from? The paper's third result links it to the hippocampal memory system. During sharp-wave ripples (SWRs) — brief offline CA1 population bursts implicated in memory consolidation — mean V1 firing barely budges. But ripple-by-ripple, the identity of the active CA1 ensemble predicts which V1 neurons fire: a Poisson GLM from the CA1 population vector to V1 spike counts in the 0–200 ms post-ripple window significantly predicted 55% of V1 neurons with appreciable ripple activity, and outperformed a scalar "mean CA1 activity" control. Crucially, the predictable V1 population was 87% dark-active, and among dark-active cells, the SWR-predictable ones had significantly higher path-invariance. So the neurons carrying the internal code are precisely the ones woven into hippocampal-cortical ensemble coordination — circumstantial but suggestive evidence that the scaffold is learned and maintained through this network, consistent with prior reports of coordinated V1–hippocampal replay (Ji & Wilson 2007).

What changes if this holds — and what to doubt

If it holds, the framing shift is real: primary sensory cortex is not a feed-forward feature bank that cognition tweaks; for most of its neurons (73% dark-active here), it is a substrate for an internal model of task state, with sensory input acting as a multiplicative correction term. That is the predictive-processing / world-model picture made concrete at the level of single-neuron gain arithmetic, and it dovetails with abstract "task progress" codes recently reported in prefrontal and parietal cortex (Guidera et al., El-Gaby et al., Johnson et al.). It also suggests a general architectural motif — shared low-dimensional task scaffold, context-specific gains — that other sensory cortices may reuse.

Now the skepticism. The central vulnerability is the word "cognitive." Same-turn paths share, by construction, a similar motor program, and the GLM motor model — six kinematic variables from overhead tracking — cannot capture everything an animal does: eye movements, whisking, pupil, posture, or preparatory/efference-copy signals timed to progress through the turn sequence. V1 is demonstrably driven by head movements and vestibular input (some of it from this same lab). The authors show measured kinematics are near-identical across light and dark (fig. S5) and that progression adds predictive power beyond measured motor variables, but an unmeasured motor-plan signal locked to the turn sequence would be observationally identical to a "cognitive" progression code in this design. Indeed, the poor transfer to opposite-turn paths is exactly what a mirrored motor program predicts. What the data cleanly establish is an internally generated, turn-sequence-locked, vision-independent code; whether it's an abstract task variable or a rich sensorimotor plan is a distinction this experiment can't fully close — though one might reasonably ask whether that distinction is even well-posed.

Other caveats: four male rats, well-trained only (>90% correct, days 5–7), so nothing here about how the code emerges with learning — a shame, since the SWR link begs for a learning timecourse. Effect sizes in bits/spike are modest (medians 0.03–0.08 for the model comparisons), even if the sign is remarkably consistent across neurons and animals. The multiplicative model won for 54% of neurons in the three-way test — a plurality, not a monolith; a quarter of neurons still look more like independent visual responders. The SWR analysis shows correlation with CA1 ensembles, not a causal or directional pathway. And it's a preprint.

Where to spend your time

The stimulus-swap analysis (Fig. 3 and its methods section, "Models relating dark and light activity") is the most valuable part — it's a genuinely clever use of an environmental manipulation to arbitrate between gain, addition, and replacement, and the kind of design worth stealing for other modulation questions. Fig. 1D–G carries the core path-invariance evidence; fig. S5 is where to go to stress-test the motor confound yourself. Everything — raw NWB data on DANDI, analysis code, even a pre-populated database in a Docker image — is public, which for a claim this contrarian is exactly how it should be.