Rapid alternation of conscious and suppressed stimulus representations in the primate brain

TL;DR

Hesse et al. use Neuropixels recordings across inferotemporal (IT) and prefrontal cortex in macaques during binocular flash suppression — a paradigm where two different images are shown to the left and right eyes and only one reaches awareness — and find that neural activity rapidly alternates (~15 Hz, ~60 ms micro-epochs) between a “conscious” mode and a “suppressed” mode. During each micro-epoch, a pure signal of the respective percept dominates: from the conscious epochs alone, the consciously perceived image can be accurately reconstructed from population activity, and from the suppressed epochs, the suppressed image can be. Single neurons concertedly switch their tuning across these micro-epochs, with feature tuning amplified during conscious moments. The claim: conscious access is not a matter of where in the brain the signal lives but when — a temporal selection on top of spatial localization.

Key claims

• During a constant stimulus and percept, neural activity in IT and prefrontal cortex rapidly alternates between two coordinated states ~15 times per second [§Abstract].
• Reconstructing images from population activity restricted to one of the two states yields the consciously perceived image; restricting to the other state yields the suppressed image [§Abstract].
• Single neurons switch tuning in sync with these micro-epochs — representing the conscious stimulus during conscious epochs and the suppressed stimulus during suppressed epochs [§Abstract].
• The two epochs are not symmetric: feature tuning is amplified during conscious micro-epochs, providing a reliable signature distinguishing them [§Abstract].
• Prefrontal cortex recordings yield qualitatively similar results, placing the alternation across both high-level visual and frontal regions [§Abstract].
• Methodology: binocular flash suppression (controllably dissociates stimulus from percept) + Neuropixels recordings across multiple face patches, thousands of neurons, thousands of stimulus pairs [§Abstract].

Method

Macaques perform binocular flash suppression: one image is presented to one eye, and a flashed image to the other eye renders the first image perceptually suppressed. This gives experimental control over which of two physically present stimuli reaches awareness. The authors record from thousands of neurons across multiple face patches in IT cortex and in prefrontal cortex using Neuropixels probes. Within the response to each stimulus pair, they identify two latent neural states (micro-epochs of ~60 ms each) that alternate at ~15 Hz. State assignment is population-level and unsupervised with respect to the behavioural report. For each state, they fit a decoder that reconstructs the stimulus image from population activity; the resulting reconstructions sharply align with the conscious image for one state and the suppressed image for the other. Single-neuron analyses show that the same cells carry both representations at different moments, with the conscious micro-epoch amplifying feature tuning.

Results

• Alternation rate: ~15 Hz (~60 ms per micro-epoch) [§Abstract].
• Decoded reconstructions from the “conscious” micro-epochs look like the consciously perceived image; reconstructions from the “suppressed” micro-epochs look like the suppressed image — a result the authors describe as a pure correlate of conscious awareness, not a weak modulation on top of a stimulus-driven signal (which is what prior textbook accounts had reported) [§Abstract].
• Replication of the IT finding in prefrontal cortex [§Abstract].
• The authors frame this as resolving a decade-long puzzle: prior literature consistently reported only weak modulation by consciousness in high-level visual cortex, which contradicted the vivid subjective quality of rivalry. Averaging across time was washing out the alternation [tweet thread, 9–11/].

Why it’s interesting

This is outside the usual Luma reading list (video generation, world models, robotics, LLM architecture), but it’s a Doris Tsao paper on representations in the primate visual system and it bears on how we think about representation geometry in biological networks — the closest wiki neighbour is A foundation model of vision, audition, and language for in-silico neuroscience (TRIBE v2) (Meta’s TRIBE v2 brain-encoding foundation model), which assumes a (mostly) stationary mapping from stimulus to BOLD per subject; if IT activity actually alternates between percept states at 15 Hz, then any model that regresses the time-averaged BOLD against a stimulus feature is averaging across this alternation, which could explain part of the headroom brain-encoding models leave on the table at fine temporal scales (though fMRI’s own TR is far coarser than 60 ms, so the practical impact for TRIBE-style work is indirect). It’s also a counterexample to the “where does X live in the network” framing that dominates Mechanistic Interpretability in artificial networks: the signal here isn’t localized to a region but to a temporal phase of a shared population — a reminder that spatial decomposition can miss structure that only shows up when you cut the data along time. Finally, the “same neurons, different micro-epochs, different content” phenomenology rhymes loosely with superposition arguments in SAE/mech-interp work, where a single unit multiplexes many features — here the multiplexing is explicit in time.

See also

• A foundation model of vision, audition, and language for in-silico neuroscience (TRIBE v2) — Meta TRIBE v2, brain-encoding foundation model that predicts fMRI from vision/audio/language; this result suggests any fine-timescale encoding model needs to account for ~15 Hz state alternation
• Mechanistic Interpretability — the finding is a biological analogue of “the right decomposition is not always spatial”