henhaobio
Sign in

bioRxivNeurosciencePreprint

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

Janis Karan HesseiD, Frank F Lanfranchi, Conor McGrory, Yuelin ShiiD, Tatiana A. EngeliD, Doris Ying TsaoiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

The neural correlates of consciousness remain elusive [1]. Previous studies found that high-level visual cortex is weakly modulated by conscious perception [2,3], yet a pure signal of conscious perception has not been found anywhere in the brain. Here, we employed Neuropixels recordings across multiple face patches in inferotemporal and prefrontal cortex during a binocular flash suppression paradigm [4], in which one of two images is consciously perceived while the other, though continuously present, is suppressed from awareness. We found that, during a constant stimulus and percept, the brain rapidly alternates, about 15 times per second, between conscious and suppressed micro-epochs. During conscious micro-epochs, a pure signal of conscious perception was present, allowing accurate reconstruction of the conscious image from neural activity. During suppressed micro-epochs this reversed: the pure signal now tracked the suppressed stimulus, allowing accurate reconstruction of the suppressed image. At the single-cell level, neurons responded selectively to the conscious stimulus and not to the suppressed stimulus during conscious micro-epochs; conversely, during suppressed micro-epochs, neurons responded selectively to the suppressed stimulus. Despite this apparent symmetry, the two epochs could be reliably distinguished, with feature tuning amplified during conscious micro-epochs. Recordings in prefrontal cortex yielded qualitatively similar results. Together, these findings show that both inferotemporal and prefrontal cortex rapidly interleave conscious with suppressed content, indicating that conscious access requires temporal selection beyond localization.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (16 Oct 2026) · CC BY 4.0

With stimulus and percept fixed, macaque inferotemporal face-patch populations alternate about 15 times per second between states purely representing the conscious or suppressed face; prefrontal cortex shows qualitatively similar interleaving.

Rivalry studies found neurons only weakly modulated by the conscious percept relative to unambiguous presentation and still informative about the suppressed image, leaving no pure correlate of conscious perception. Simultaneous Neuropixels recordings resolve single-trial population states, parametric faces let those states be reconstructed, and the weak modulation is explained by averaging over interleaved conscious and suppressed micro-epochs plus slightly stronger conscious-state tuning.

Key results

  1. ShownWithin a constant percept, inferred population states each carry a pure representation. Activity in state 1 reconstructs the consciously perceived face, and activity in state 2 reconstructs the suppressed face. In each state, information about the other face is near chance.

    State 1 reached over 80% explained variance for some conscious-face features (priming upper bound over 90%), with the suppressed face near chance; the pattern reversed in state 2 (p = 3 × 10^−29, n = 30 pairs). Held-out-half validation: p = 6 × 10^−17. Single-trial 10-way decoding gave over 61% conscious vs 4% suppressed in state 1, against 40% and 34% from time-averaged activity. · Reconstructions from inferred states recover…; Coordinated population states alternate between…; Figure legends

  2. ShownThe alternation is fast and coordinated across the population, with state dwell times near 60 ms. Simulated models in which responses are averaged, or in which each neuron switches independently, do not reproduce it.

    LLR autocorrelation time constant 67 ms (shuffled data decay within one bin); HMM median dwell 64 ms (state 1) and 59 ms (state 2) at 5 ms bins. The two-state model beat the best one-state model in all 19 sessions (p = 10^−4). Of four simulated models, only the two-state model passed all four tests; the averaging and uncoordinated models failed held-out reconstruction, reversal and likelihood. · Single-trial population activity can be…; Coordinated population states alternate between…; Dynamics of state switching; A battery of tests

  3. ShownTemporal multiplexing explains the weak perceptual modulation seen in rivalry. Averaged over time, modulation is much weaker than with unambiguous stimuli. Within state 1 it matches unambiguous selectivity, and within state 2 it inverts.

    Time-averaged binocular vs unambiguous modulation: slope 0.23 (p = 8 × 10^−133, 5550 cell-stimulus pairs). In held-out neurons (2885 pairs): state 1 slope 0.93 (r = 0.66), state 2 slope −0.76 (r = −0.44). After equalizing state gains, the residual correlation vanished (r = −0.02, p = 0.31; slope 0.07). · Average neural activity is only weakly modulated…; Temporal multiplexing explains weak average…; Figure legends

  4. ShownThe same neurons encode both faces, but tuning gain for the conscious face in state 1 exceeds gain for the suppressed face in state 2, and mean firing is slightly higher in state 2. Because of these differences, a linear decoder can tell the two states apart despite their apparent symmetry.

    Across 543 cells, the two gains were correlated (r = 0.90) and conscious-state gain was higher (p = 3 × 10^−40; mean ratio 1.21). Mean firing was slightly higher in state 2 (p = 2 × 10^−15). A linear decoder separated state 1 from state 2 responses with 66% accuracy. The state 2 suppressed-face tuning axis matched the priming axis relative to noise ceiling (p = 0.35). · Single cells have higher gain of tuning during…; Figure legends

  5. SuggestedOrbitofrontal and ventrolateral prefrontal populations show qualitatively similar interleaving, though reconstruction quality is lower. Time-averaged prefrontal activity likewise lacks a pure conscious-percept signal.

    Priming upper bound below 80% explained variance (shape features under 30%) vs over 90% in IT. Held-out-half separation p = 3 × 10^−4 (30 pairs); held-out decoding p = 0.03. Modulation slope 0.32 for averaged activity vs 0.91 in state 1 and −0.52 in state 2 (283 cell-stimulus pairs); after gain equalization r = 0.24 (p = 0.005). · Evidence for similar interleaved representations…; Figure legends

Methods, data and reproducibility

Neuropixels NHP probes recorded inferotemporal face patches ML and AM and prefrontal regions PO and PVL in two macaques during binocular flash suppression with up to 1500 parametric faces. A Poisson log-likelihood ratio decoder or two-state HMM with priming-derived templates labelled time bins as conscious or suppressed; PCA plus linear regression trained on priming responses reconstructed face features.

  • Neuropixels recordings from ML and AM during flash suppression of 1500 parametric faces (750 pairs; 60 high-repeat faces used for reconstruction) · rhesus macaque · n = 9 sessions (5 in monkey A, 4 in monkey M)
  • Neuropixels recordings during flash suppression of 25 pairs built from 10 screened stimuli, including an unambiguous alternation condition · rhesus macaque · n = 10 sessions (7 in monkey A, 3 in monkey M)
  • Visually responsive neurons pooled across sessions for face-feature reconstruction · rhesus macaque · n = 3501 neurons
  • Prefrontal recordings from orbitofrontal (PO) and ventrolateral (PVL) regions · rhesus macaque · n = 661 prefrontal neurons
  • Human psychophysics: button-press percept reports for the same flash suppression displays · human · n = 14 participants
  • fMRI faces-versus-objects localizer (3T, MION contrast agent) used to target face patches · rhesus macaque · n = 2 male rhesus macaques
Methods & tools
Neuropixels 1.0 NHP probes, SpikeGLX, Kilosort3, hmmlearn, Python 3.11, MATLAB, Poisson hidden Markov model (Baum-Welch, Viterbi), linear support-vector machine, active appearance model, fMRI with MION contrast, Eyelink, Claude Code
Reagents & environment
Python 3.11 for LLR and HMM fitting (HMM via hmmlearn); other analyses in MATLAB 2023b and 2025b; acquisition with SpikeGLX and spike sorting with Kilosort3.
Reproducibility gaps
Electrophysiology data are to be deposited in the DANDI Archive and custom code on GitHub only upon publication; no links or accessions are given.
Limitations (6)
  • State-conditioned averages that reuse the same spikes for state assignment and evaluation are inflated by circular conditioning. The authors therefore rely on held-out-neuron validation, which gives modestly lower accuracy. · stated by the authors
  • Micro-epochs were only weakly coordinated between simultaneously recorded ML and AM. Cross-patch reconstructions were low, and for AM to ML the conscious face exceeded the suppressed face in state 2. · stated by the authors
  • Prefrontal reconstructions were weaker (priming upper bound below 80%, shape features below 30%), possibly reflecting fewer face cells or a different feature code there. · stated by the authors
  • A discrete two-state model forces gradual biological transitions into one of two states. This may lower state-specific reconstruction accuracy relative to the idealized simulation. · stated by the authors
  • No perceptual report from the monkeys is described. Conscious and suppressed labels rest on reports from 14 human participants viewing the same displays, not on per-trial measurements in the animals. · noted in this summary
  • Only two monkeys were recorded. State-resolved reconstructions were scored on 60 high-repeat faces (30 stimulus pairs), and the 1500-face experiment had no unambiguous alternation condition. · noted in this summary

Read this if you study neural correlates of consciousness or binocular rivalry, or need methods for inferring fast single-trial population states from large-scale recordings while controlling for circular-conditioning bias.

Citations
2

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKapoor '20Hesse '20Fang '24Mai '19Dwarakanath '23Weilnhammer '21
79% semantic matchNeuroscience
Vishal Kapoor, Abhilash Dwarakanath, Shervin Safavi+4 more·bioRxiv·Jan 28, 2020· 48 citations