Contribution
Most searches for the neural correlates of consciousness look at onsets or at seen-versus-unseen contrasts with degraded stimuli. This paper varies how long a clearly seen image stays on, with no report on the analysed trials, and follows the content of activity over time and not only its magnitude. It shows that sustained content lives in a stable population code in ventral visual cortex while local activity decays, and that prefrontal content is a transient onset burst.
Key insight
Amplitude and content come apart. A region can quiet down by four fifths while keeping a pattern that a fixed readout decodes as well at the end as at the start. So “the response faded” does not mean “the representation faded”. What lasts as long as the percept is a fixed direction in population space, which the authors call an “experience subspace”, and not the level of firing.
Assumptions
- Participants: 10 patients with intractable epilepsy (4 female, age 19–65), 64–128 subdural electrodes each, 8 of 10 right hemisphere only; electrode placement clinical.
- Stimuli and task: greyscale images about 5° in size, of faces (about 30%), watches (about 30%), other objects, animals and rare targets; durations 300–1,500 ms. Patients pressed a key only for targets (clothing, or a late blur in a dual task, together about 10% of trials); target trials were not analysed.
- Signal: HFA, 70–150 Hz, the mean of eight normalised 10 Hz Hilbert-amplitude bands, baseline-corrected and smoothed over 50 ms, taken as a proxy for local firing.
- Electrode selection: responsive if HFA differs from baseline in any of four 200 ms windows for any category (Bonferroni across windows, FDR across electrodes), without regard to category or timing: 430 of 907 noise-free electrodes.
- Pooling: the main multivariate analyses pool electrodes across patients into one pseudo-population per region, with single-patient checks in the supplement.
- Decoding: regularised LDA per time point, five-fold cross-validation repeated five times, AUC, cluster-based and max-statistic permutation tests; 92 images seen by all patients for 900 ms or more.
Key results
- Attenuation (Fig. 1). Mean attenuation of HFA from peak to 800–900 ms is 81.8% ± 1.1% across Occ, VT, Par and PFC; category selectivity (η²) falls 77.4% ± 1.1%. Of 430 responsive electrodes, 236 are category selective.
- Trajectories (Fig. 2). Multivariate distance from baseline falls by about 80%, but in VT and Occ it separates by duration shortly after the shorter stimulus ends; PFC and Par trajectories do not.
- Category decoding (Fig. 3). Peak AUC: VT 99.6%, Occ 98.1%, PFC 84.1%, Par 86%. Face–watch AUC averaged over 100–900 ms: VT 99.8%, Occ 92.4%. In PFC and Par significant decoding is mostly 150–600 ms. The temporal generalisation matrix is square in VT (time-invariant code), not in PFC or Par. Controls for ocular-muscle artefacts (on OFC electrodes), using all PFC electrodes rather than responsive ones, and subsampling VT and Occ to PFC electrode counts leave the pattern unchanged. Category information is stronger in OFC than in LPFC.
- Duration (Fig. 4). VT information outlasts offset by about 450 ms for 300 ms images and 230 ms for 900 and 1,500 ms images; 900−300 and 1,500−900 contrasts are significant in VT and Occ (cluster p < 0.006) and not in PFC or Par (p > 0.2).
- Exemplars (Fig. 5). On 60 images seen at least twice by 5 patients, item and geometry reliability are sustained and stable in VT through 900 ms, also after partialling out category models; less robust in Occ; transient and weaker in PFC and Par (IR significant in both; GR only in Par, not after removing category structure).
- Theory predictions. IIT (stable, duration-long posterior representation) and GNWT (transient prefrontal onset ignition without a sustained component, absent report) are both met. GNWT’s predicted offset ignition was not seen.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | In ventral temporal cortex, category and exemplar content of a seen image is represented in a time-invariant population code for as long as the image is on, while amplitude falls about 80% | strong for VT categories; moderate for exemplars (5 patients, 60 images); weaker in Occ | Figs. 1, 3–5 |
| C2 | Prefrontal and parietal cortex carry category and exemplar content transiently after onset, without report and not tracking duration | moderate: few PFC electrodes, mostly one hemisphere; absence of sustained PFC coding is a null result | Figs. 3–5; Limitations |
| C3 | The multi-duration predictions of IIT and GNWT for this paradigm are both confirmed, and are not adversarial | moderate as stated; GNWT’s offset ignition was not found | Discussion |
| C4 | Sustained experience may rely on sensory representation and discrete updating on frontoparietal representation | weak: a conditional interpretation | Discussion, Conclusion |
| C5 | Perceptual quality is fixed by the population’s position in a stable “experience subspace”, not by response amplitude | speculative: a proposal | Discussion |
Concepts
- HFA (high-frequency activity): 70–150 Hz broadband amplitude, a proxy for local firing.
- temporal generalisation matrix (TGM): decoding accuracy for every pair of training and testing times; a square block means a stable code.
- item reliability (IR) / geometry reliability (GR): the paper’s repetition-based measures of whether each exemplar’s position, and the whole dissimilarity structure, recur across presentations.
- experience subspace: the authors’ name for the stable subspace of population activity that keeps category and exemplar distinctions while other dimensions vary.
- no-report: analysed trials need no response; here they were still task-relevant, since each image had to be judged a non-target.
Connections
- Dennett & Kinsbourne 1992 (LIT-431, NOTE-367). Cited as the paper’s ref. 2, for colour phi and postdiction. The paper sets discrete against continuous perception; NOTE-367 records Dennett and Kinsbourne’s argument that the Orwellian and Stalinesque readings of colour phi cannot be told apart, which this paradigm does not address.
- Oizumi et al. 2014 (LIT-445). IIT 3.0. The paper tests a COGITATE prediction attributed to IIT (a sustained posterior hot zone), not Φ or any postulate.
- Butlin et al. 2023 (LIT-056). Its GWT indicators derive from the workspace theory whose ignition prediction holds here.
- Verzhbinsky et al. 2026 (LIT-633). Also human intracranial high-frequency signals, but about co-occurrence across regions, not representation within them.
Bearing on the record
- THEORY-023. It holds that theory-derived indicators presuppose the disputed theory. This reading shows two rival theories making compatible predictions in one paradigm; the predictions were tested, but the theories were not discriminated. It neither supports nor contradicts THEORY-023’s claim about AI.
- A candidate THEORY: in human visual cortex the content of a sustained percept is carried by a stable population code, not by sustained activity level. One source; a second (Podvalny et al. 2017, cited here, not in the record) would strengthen it.
- No ML instruction. Nothing here belongs in the anthology.
Limitations
- Sparse and lateralised PFC coverage; the authors warn that the absence of sustained prefrontal activity should be taken “with more caution than the presence of activity”.
- Only seen stimuli. Without unseen or rivalrous stimuli, nothing shows that the sustained VT code is specific to experience; the authors propose binocular rivalry as the next test.
- One frequency band. PFC could hold the percept in low-frequency activity or silent synaptic states, which HFA does not measure.
- Pooled pseudo-populations across patients for the main analyses.
- “Representation” is correlational in the authors’ own definition: patterns that correlate with stimulus descriptors, with no mechanistic role implied.
- Task relevance. Analysed trials needed no response but needed a non-target judgement, so they are not task-free.
Open questions
- Does the VT code track the percept when percept and input come apart, as in rivalry, or when an image is present but unseen?
- Is there a sustained prefrontal representation in bands or sites not sampled?
- How does an “experience subspace” relate to representational drift over longer timescales, which the authors raise and leave open?
Corrections
- none to a seeded skim (there was no seed)