The predictive global neuronal workspace: A formal active inference model of visual consciousness

The predictive global neuronal workspace: A formal active inference model of visual consciousness Whyte Smith

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

The predictive global neuronal workspace: A formal active inference model of visual consciousness 2020 Progress in Neurobiology 10.1016/j.pneurobio.2020.101918 Whyte2020-the-predictive-global-neuronal-workspace-a-f Whyte2020-the-predictive-global-neuronal-workspace-a-f.txt, the project's staged copy b8f5a54b6c25808891312316629d7f3e65a96a77 not-redistributable neuroscience

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them mechanism Quotation withheld because this record is not cleared for redistribution. visual consciousness (conscious access) Quotation withheld because this record is not cleared for redistribution. simulation simulation Quotation withheld because this record is not cleared for redistribution. mechanisms Conscious access happens because a temporally deep level of the hierarchy infers and integrates lower-level states over a slow enough timescale to contextualise them and drive extended report behaviour. Quotation withheld because this record is not cleared for redistribution. mechanisms Late ERPs arise because the ERP is the time derivative of the depolarisation variable, so a rapid change in high-level belief produces a large late deflection. Quotation withheld because this record is not cleared for redistribution. mechanisms Rapid belief change increases the magnitude of the depolarisation derivative, which is what makes a late component appear. Quotation withheld because this record is not cleared for redistribution. mechanisms Expectations act by raising second-level posterior confidence, which magnifies top-down feedback to sensory populations and pushes marginal trials past the ignition threshold. Quotation withheld because this record is not cleared for redistribution. Scope deliberately restricted to settings where visual consciousness is operationalised by report. what this limits: every prediction the model makes: WE HAVE DELIBERATELY LIMITED THE SCOPE OF THE MODEL TO EXPERIMENTAL SETTINGS WHERE VISUAL CONSCIOUSNESS IS OPERATIONALISED VIA REPORT Quotation withheld because this record is not cleared for redistribution. Report paradigms themselves carry known methodological problems. what this limits: the same scope as limitations[0], with the cost: REPORT PARADIGMS COME WITH METHODOLOGICAL DIFFICULTIES Quotation withheld because this record is not cleared for redistribution. Restricting the model to vision perpetuates the visual bias of consciousness science. what this limits: the modality the theory has been tested in, and what that does to the field: IN LIMITING THE SCOPE OF THE PGNW TO THE VISUAL MODALITY, WE ARE AWARE THAT WE ARE REIFYING THE PERVASIVE BIAS IN CONSCIOUSNESS SCIENCE OF PRIMARILY STUDYING VISION Quotation withheld because this record is not cleared for redistribution. The match to the four-way taxonomy is partly built in, since parameters were tuned to the published behavioural results. what this limits: the model's match to the four-way taxonomy, which is not independent confirmation: THIS SHOULD NOT BE A SURPRISE, AS WE FINETUNED THE PARAMETER VALUES TO CAPTURE BEHAVIOUR REPORTED IN THE EMPIRICAL LITERATURE Quotation withheld because this record is not cleared for redistribution. The model cannot represent partial awareness, because information cannot skip hierarchical levels in it. what this limits: any account of partial awareness: THIS REQUIRES INFORMATION TO SKIP LEVELS OF THE HIERARCHY, WHICH IS ALSO NOT POSSIBLE IN THE PRESENT MODEL Quotation withheld because this record is not cleared for redistribution. Simulated report frequencies are only directional, not quantitative, predictions. what this limits: the simulated report frequencies: THEY SHOULD BE INTERPRETED AS DIRECTIONAL HYPOTHESES, AS OPPOSED TO PRECISE PREDICTIONS ABOUT THE PERCENTAGE OF SEEN VS UNSEEN TRIALS Quotation withheld because this record is not cleared for redistribution. Two predictions of the theory this paper extends are stated to have been falsified. what this limits: any appeal to the original global neuronal workspace formulation: the authors state that as experimental paradigms became more sophisticated, two of its predictions have been falsified Quotation withheld because this record is not cleared for redistribution. The falsifying result is qualified inline by whether its instrument is a fair proxy. what this limits: the claim that information is unconsciously maintained by the same structures: the authors qualify it as holding Quotation withheld because this record is not cleared for redistribution. The P3 is reported to track task relevance rather than conscious access. what this limits: any use of the P3 as a signature of conscious access: a masking-paradigm replication is stated to show unequivocally that Quotation withheld Quotation withheld because this record is not cleared for redistribution. A marker was abandoned without any accompanying theoretical revision, which the authors name as a degenerative pattern. what this limits: any reading of the GNW's response to the P3 findings Quotation withheld because this record is not cleared for redistribution. The replacement model offered for the falsified one is stated to be too idealised for neurobiological predictions. what this limits: any neurobiological prediction attributed to the revised computational model of subjective report: the authors state it is too idealised to make them Quotation withheld because this record is not cleared for redistribution. Where phenomenology sits in the model was chosen to avoid an empirically unfalsifiable consequence. what this limits: the model's placement of phenomenology: the alternative would make the presence of a particular phenomenology empirically unfalsifiable, and the authors state they situate it as they do to avoid that conclusion Quotation withheld because this record is not cleared for redistribution. A parameter that empirical work fits to data was fixed by assumption, and it affects the predicted ERPs. what this limits: any ERP predicted by these simulations: the decay parameter that in empirical work needs to be fit to neuronal or behavioural responses was set to 1 here, and the authors state predicted ERPs can be affected by different values of it Quotation withheld because this record is not cleared for redistribution. the Active Inference formalism used to cast global neuronal workspace dynamics as approximate Bayesian inference, yielding testable behavioural and neural predictions Quotation withheld because this record is not cleared for redistribution. a debrief after phase one measured how many participants had noticed the task-irrelevant square at all eliminates: that the stimulus was consciously seen - the inattentional-blindness manipulation is checked rather than assumed, and 50 per cent reported not seeing it · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. P3 P3 Quotation withheld because this record is not cleared for redistribution. conscious access conscious access Quotation withheld because this record is not cleared for redistribution. task relevance task relevance Quotation withheld because this record is not cleared for redistribution. multivariate decoding multivariate decoding Quotation withheld because this record is not cleared for redistribution. degenerative research programs degenerative research programs Quotation withheld because this record is not cleared for redistribution. phenomenology phenomenology Quotation withheld because this record is not cleared for redistribution. mdp.erp mdp.erp Quotation withheld because this record is not cleared for redistribution. the P3 - A MARKER WITHDRAWN WITHOUT A REVISION: the claim that it is a specific marker of conscious access is no longer defended, but the abandonment was not accompanied by any theoretical revisions and makes no additional predictions, which the authors call a recognized characteristic of degenerative research programs Quotation withheld because this record is not cleared for redistribution. the falsifying evidence - CONDITIONAL ON A PROXY: information is said to be unconsciously maintained by the same structures at least briefly and insofar as multivariate decoding is a defensible proxy for information processing, so the result that overturns the theory is bounded by an instrument assumption stated in its own sentence Quotation withheld because this record is not cleared for redistribution. phenomenology, in this model - SITED TO KEEP IT FALSIFIABLE: the alternative placement would make the presence of a particular phenomenology empirically unfalsifiable, and the authors situate it at the point where lower-level representations update higher-level states to avoid that Quotation withheld because this record is not cleared for redistribution. the consistent parameter set - FITTED, WITH THE AUTHORS' ARGUMENT FOR WHY IT STILL COUNTS: the values were finetuned to capture behaviour reported in the literature, and the authors argue that no consistent set need have existed, so finding one supports the model Quotation withheld because this record is not cleared for redistribution. extends → the global neuronal workspace (GNW) model Quotation withheld because this record is not cleared for redistribution. builds-on → Hohwy (2013), Whyte (2019), and Friston et al (2012) Quotation withheld because this record is not cleared for redistribution. contradicts → Fleming's (2019) higher-order state-space (HOSS) model Quotation withheld because this record is not cleared for redistribution. P3 —distinct-from→ conscious access P3 conscious access strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. task relevance —contrasted-with→ conscious access task relevance conscious access differs in: which of the two the P3 tracks · strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. variational Bayesian inference variational Bayesian inference Quotation withheld because this record is not cleared for redistribution. parameter fitting (grid search) parameter fitting (grid search) Quotation withheld because this record is not cleared for redistribution. With one consistent parameter set the simulated agent reproduced the GNW four-way taxonomy behaviourally: weak signal plus absent attention gave chance forced-choice performance and almost no 'seen' reports. Quotation withheld because this record is not cleared for redistribution. Simulated first-level ERPs, corresponding to early components such as P100 and N100, barely changed with visibility. Quotation withheld because this record is not cleared for redistribution. Simulated second-level ERPs, corresponding to the late P3, were strongly modulated by visibility, matching the empirical early-versus-late ERP dissociation. Quotation withheld because this record is not cleared for redistribution. The model reproduced the Pitts inattentional-blindness dissociation: only in phase three, when the square was both seen and task-relevant, did a large late P3-like ERP appear. Quotation withheld because this record is not cleared for redistribution. The model predicts that visual consciousness will be accompanied by a late ERP whenever high-level confidence changes rapidly. Quotation withheld because this record is not cleared for redistribution. Adding prior expectation as a third factor, expectations consistent with the stimulus raised both forced-choice accuracy and the rate of 'seen' reports in every quadrant of the taxonomy. Quotation withheld because this record is not cleared for redistribution. With a strong attended stimulus, a consistent prior shrank the simulated P3 while an inconsistent prior enlarged it relative to neutral priors. Quotation withheld because this record is not cleared for redistribution. The model reconciles univariate ignition results with multivariate decoding of unseen stimuli, because above-baseline firing persists on unconscious trials and would be readable by a classifier even though mean rates are lower. Quotation withheld because this record is not cleared for redistribution. global workspace theory (GWT) global workspace theory (GWT) Quotation withheld because this record is not cleared for redistribution. global neuronal workspace (GNW) model global neuronal workspace (GNW) model Quotation withheld because this record is not cleared for redistribution. ignition ignition Quotation withheld because this record is not cleared for redistribution. access consciousness access consciousness Quotation withheld because this record is not cleared for redistribution. predictive global neuronal workspace (PGNW) predictive global neuronal workspace (PGNW) Quotation withheld because this record is not cleared for redistribution. Active Inference Active Inference Quotation withheld because this record is not cleared for redistribution. partially observable Markov decision process (POMDP) partially observable Markov decision process (POMDP) Quotation withheld because this record is not cleared for redistribution. attention attention Quotation withheld because this record is not cleared for redistribution. stimulus signal strength stimulus signal strength Quotation withheld because this record is not cleared for redistribution. prior probability prior probability Quotation withheld because this record is not cleared for redistribution. forced choice performance forced choice performance Quotation withheld because this record is not cleared for redistribution. firing rate firing rate Quotation withheld because this record is not cleared for redistribution. state transition state transition Quotation withheld because this record is not cleared for redistribution. 200 trials — size of the proof-of-principle simulation, half with the square present Quotation withheld because this record is not cleared for redistribution. 100% — forced-choice and report accuracy of the model under full attention and high signal strength Quotation withheld because this record is not cleared for redistribution. 47% — chance-level forced-choice accuracy in the weak-signal, attention-absent quadrant Quotation withheld because this record is not cleared for redistribution. 58% and 61% — forced-choice accuracy when signal strength alone, or attention alone, was raised Quotation withheld because this record is not cleared for redistribution. 53% — rate at which the calibrated model reported seeing the square in phase one of the simulated inattentional-blindness task, matching the human 50% Quotation withheld because this record is not cleared for redistribution. 800 — square-present trials simulated for the expectation-extended taxonomy, 100 per prior-by-quadrant cell Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1016/j.pneurobio.2020.101918.