Decoding internally generated transitions of conscious contents in the prefrontal cortex without subjective reports

Abstract

A major debate about the neural correlates of conscious perception concerns its cortical organization, namely, whether it includes the prefrontal cortex (PFC), which mediates executive functions, or it is constrained within posterior cortices. It has been suggested that PFC activity during paradigms investigating conscious perception is conflated with post-perceptual processes associated with reporting the contents of consciousness or feedforward signals originating from exogenous stimulus manipulations and relayed via posterior cortical areas. We addressed this debate by simultaneously probing neuronal populations in the rhesus macaque (Macaca mulatta) PFC during a no-report paradigm, capable of instigating internally generated transitions in conscious perception, without changes in visual stimulation. We find that feature-selective prefrontal neurons are modulated concomitantly with subjective perception and perceptual suppression of their preferred stimulus during both externally induced and internally generated changes in conscious perception. Importantly, this enables reliable single-trial, population decoding of conscious contents. Control experiments confirm significant decoding of stimulus contents, even when oculomotor responses, used for inferring perception, are suppressed. These findings suggest that internally generated changes in the contents of conscious visual perception are reliably reflected within the activity of prefrontal populations in the absence of volitional reports or changes in sensory input.

Fig. 1. Recorded area, BR paradigm and example unit activity.

a, Utah array location displayed over the vlPFC (schematic macaque brain and post implantation location in one animal). b, Visual input, OKN and spiking activity during an example BR and PA trial. A drifting sinusoidal grating (downward) was monocularly presented first during both trials. During BR, an upward drifting grating was presented to the contralateral eye 2000 ms later, resulting in perceptual suppression (BFS) of downward motion, as inferred from the OKN (grey curve). Externally induced perceptual suppression lasted ~3000 ms, following which a spontaneous switch (~5000 ms) reinstated the perception of downward motion (green curve). Units 33 and 119 display strong spiking activity when downward drifting grating is perceptually dominant, while units 44 and 167 respond more when upward drifting grating is perceived. Below is a PA trial. Following initial monocular presentation of downward drifting grating, an upward drifting grating was presented to the contralateral eye. The stimulus was switched later (~4500 ms) resulting in a change in OKN polarity. Individual units were preferentially modulated by similar direction of motion as during BR c, Perceptual dominance distributions during flash suppression and rivalry phases (derived from the OKN traces) are approximated well by a gamma distribution. d, Sites with significant stimulus preference (Wilcoxon rank-sum test, two-sided, p ≤  0.05) during PA trials (physical alternation phase) projected on the array for one recording session (selectivity here was computed using spiking activity recorded from a given electrode). Location of units (identified after spike sorting) presented in b are marked. Green and pink pixels reflect sites, where activity was stronger for downward or upward drifting gratings, respectively. e, Average spike density functions of two simultaneously recorded PFC units, 33 (preferred downward motion) and 167 (preferred upward motion), during PA and BR trials. Pink and green colors in the first four columns depict responses pertaining to downward and upward drifting grating respectively. The last two columns display the activity during a stimulus or perceptual switch from downward to an upward drifting grating (pink) and vice versa (green). Source data are provided as a Source Data file.

Fig. 2. Sensory (PA) versus perceptual (BR) modulation of spiking activity - d’ and average population activity.

a, Scatter plot of sensory vs. perceptual selectivity (d′) for all units (dots) during BFS and BR. Displayed with different colors are units; showing no significant modulation in PA or BR trials (grey); those with significant modulation for the same stimulus during both trial types (green); units displaying significant preference only during PA (red) and units displaying significant modulation only during BR trials (blue) and a small percentage of units which fired more when their preferred stimulus was perceptually suppressed (cyan). Proportion of perceptually modulated units for both BFS and BR increased as a function of sensory selectivity strength (insets showing perceptual modulation for d’>1, BFS - 90% and BR - 86%). b, Population activity averaged across units, which were significantly modulated during PA (upper row, presented in black) or BR (lower row, presented in red) trials and preferred the same stimulus, is plotted for perceptual dominance brought about by BFS (left) or during BR (middle), as well as switches (right) during BR. Plotted above is population activity during temporally matched phases in PA. Shaded regions depict standard error of the mean. The orange dashed line indicates the average delay between the physical stimulus transition and the OKN derived transition during PA trials (129.4 ± 36.6 ms following the onset of change in the physical stimulus). A remarkable similarity in population activity across trial types indicates robust perceptual modulation. Source data are provided as a Source Data file.

Fig. 3. Decoding the contents of conscious perception from simultaneously recorded prefrontal ensembles.

a, Normalized population spiking activity of down (green) and up (pink) preferring ensembles during up to down or down to up, PA (upper row) and BR (lower row) switches shows reliable modulation. Data are presented as mean, and shaded regions depict standard error of the mean. b, Cross-temporal decoding of conscious contents around switches during PA and BR trials and generalization across the two. Classification accuracy was computed for each pair of train and test time windows (see methods) in steps of 50 ms, using 150 ms bins. In both a and b, the delay between the physical stimulus transition and the OKN derived transition during PA trials is indicated with an orange dashed line (129.4 ± 36.6 ms). c, Cross trial-type invariance of the population code assessed by training a classifier on activity during one trial type and testing on the other, before and after a switch for a single 400 ms bin (starting 200 ms pre and post switch). Box plots (for box plot description, see statistical information, methods) depict the distribution of classification results with shuffled labels (n = 500), while filled circles denote the highly significant (permutation test, one-sided, estimated p-value: p = 0.00199) classification accuracy with real labels. Results suggest invariance of the population code, thus encoding perceptual contents. The presented results were computed with data from two animals pooled together. Similar results were observed for individual animals, which are presented in supplementary figure 6. Source data are provided as a Source Data file.

Fig. 4. Assessment of decoding within individual datasets.

This figure displays the classification accuracy at the level of individual recording sessions performed across the two animals: four sessions of H07 (red asterisks) and two sessions for A11 (red circles). Each point in the scatter plot corresponds to classification accuracy during PA or BR trials, obtained in a given dataset. The three panels present the results obtained across the different temporal phases of the paradigm: a, flash dominance, b, rivalry dominance, wherein PA trials are aligned to the physical stimulus change and c, rivalry dominance, wherein PA trials were aligned to the change in OKN. The results are computed over an 800 ms window starting 200 ms after the event used for aligning the data (physical stimulus change, or OKN derived change). We note that the classification accuracy for BR in b and c are similar but not identical, because they were obtained using two separate runs of the decoding analysis. In general, we observed strong classification accuracy even at the level of individual datasets.

Fig. 5. Decoding within and generalization across recording sessions.

Displayed is the classification accuracy, a, computed for individual datasets and b, generalization accuracy across recording sessions. This analysis requires a correspondence of features across them. The units were identified after spike sorting, and can therefore be different across individual sessions. Thus, for this analysis, the spiking activity recorded from each of the 96 sites recorded with the array served as features, on which we trained and tested the classifier across the recording sessions. Results are presented for both the flash suppression (Perceptual dominance - BFS) and the rivalry phase (Perceptual dominance - BR) of the paradigm computed with neural activity elicited during an 800 ms window starting 200 ms after a stimulus or a perceptual change. The main diagonal (upper left to lower right) represents classification accuracy within a session and lateral diagonals represent generalization accuracy across sessions. While we find on average, high accuracy within individual recording sessions, the generalization across datasets is limited. Moreover, the classification accuracy within datasets was significantly more compared to generalization accuracy across datasets (two-sample t-test, two-sided, Physical Alternation_within vs. Physical Alternation_across, p = 2.42*10⁻⁹, Perceptual Dominance (BFS)_within vs. Perceptual Dominance (BFS)_across, p = 1.94*10⁻⁵, Physical Dominance_within vs. Physical Dominance_across, p = 7.45*10⁻¹⁰, Perceptual Dominance (BR)_within vs. Perceptual Dominance (BR)_across, p = 1.07*10⁻⁷). Potential reasons for such limited generalization could be that units either change or lose their preference, or it is likely, that we sampled from different units across recording sessions.

Fig. 6. Control paradigms - Fixation Off, Fixation On and example unit activity.

Trials started with a fixation spot, cueing the animal to bring and maintain gaze within a fixation window (300 ms), following which a drifting stimulus was presented monocularly. a, During Fixation Off, the fixation spot was removed at stimulus onset, thus inducing OKN. b, During Fixation On, the stimulus was presented without removal of the fixation spot, and the animal was required to maintain its gaze within a window (±1 or ±2°) until the trial ended, to receive a juice reward. c, During each trial, a stimulus drifting in one of eight different directions (pseudorandomized across trials) was presented. d, Whisker box plots (for box plot description, see statistical information, methods) displaying the distribution of standard deviations (STD) estimated from the eye movement signal (y-coordinate) elicited on individual trials during stimulus presentation (0–1000 ms). For Fixation On, either all (n = 187) or selected trials (n = 95), which displayed lower variance in the eye movement (E.M.) signal were analyzed (see methods). The STD was significantly reduced (Wilcoxon rank-sum test, two-sided, *** denotes p  ≤  0.001, Fixation Off vs. Fixation On (all trials), p = 7.86*10⁻⁶⁹, Fixation On (all trials) vs. Fixation On (low E.M. variance), p = 9.07*10⁻⁵, Fixation Off (all trials) vs. Fixation On (low E.M. variance), p = 6.09*10⁻⁴⁶) during Fixation On as compared to Fixation Off trials (n = 239). The results presented in this figure were computed with data from two animals pooled together. e and f show spike density functions overlaid on spike raster plots depicting the responses of two units to eight different motion directions during the two paradigms. The middle polar plots display the tuning curves of each unit (average response in Hz to gratings drifting in different directions). Spike rasters are displayed for first ‘n’ trials of every motion direction presentation. Here, n is the minimum number of trials presented to the animal across any motion direction during a given paradigm. PSTHs and tuning curves were computed taking all trials (of a given motion direction) into account. e, Example Unit 1 displays a stronger response to a stimulus drifting downwards during both paradigms. The unit displayed in f responds strongly to two opposite directions of motion, thus displaying orientation preference. Although the firing rate was higher during the Fixation off paradigm, the unit displayed similar preference across both paradigms. Source data are provided as a Source Data file.

Fig. 7. Decoding of motion content in the presence and during the suppression of OKN eye movements from simultaneously recorded prefrontal ensembles.

a, Ensemble population spiking activity (see methods) during fixation-Off and fixation-On paradigm for units which were significantly modulated in either paradigms and preferred the same motion direction (also see methods). Black solid and dashed lines depict the response to the preferred and nonpreferred stimulus respectively. Data are presented as mean and shaded regions depict standard error of the mean. b, Cross-temporal decoding of stimulus contents (binning parameters similar to Fig. 3b). c, Cross-paradigm invariance of the population code was assessed by training a classifier on activity during one paradigm and testing on the other, for a single 400 ms bin (starting 400 ms poststimulus onset) during visual motion presentation. Significant (permutation test, one-sided, estimated p-value: p = 0.00199) accuracy (checked by comparing it to accuracy obtained with shuffled labels (n = 500), summarized with box plots (for box plot description, see statistical information, methods)) suggests that the population code is invariant to the presence of large OKN, and encodes stimulus motion contents. Classification accuracy for decoding within the paradigm is also presented. The presented results were computed with data from two animals pooled together. We observed similar pattern of results in individual animals, and they are presented in supplementary figure 8. Source data are provided as a Source Data file.