Predictive processing and the distinction between conscious and unconscious inference
The brain predicts almost everything it senses, yet you experience only some of its predictions, and the best evidence says the difference lies in deep, slow, self-monitored inference rather than in prediction itself.
TLDR
- Claim: prediction error minimisation alone cannot separate conscious from unconscious inference. Most predictions the brain makes are useful and never experienced, so the theory needs an extra criterion it does not yet supply.
- Mechanism: the best-supported candidate is deep inference. Errors that reach slow, high-level models of the sequence, weighted by precision and tracked by a model of one’s own perceiving, go with awareness. Fast local errors do not.
- Implication: local mismatch signals survive inattention, sleep and the vegetative state, so they cannot be markers of experience. Clinical and machine tests must look for the deep, global signature instead.
- Boundary: every criterion so far was calibrated against report. A theory that defines conscious inference as reportable inference cannot tell us about experience that is present but unreported, such as the hum of a fan.
- Decision: when someone says a system “predicts”, ask at which depth, with what precision, and whether it models its own predicting. Treat prediction without those three as causally useful, not as evidence of experience.
The fan above the tabla
It is half past seven on a Thursday evening in a first-floor room in Mysuru. Deepa, a 31-year-old software tester, sits cross-legged in front of a pair of tabla. Her teacher, whom everyone calls Guruji, sits opposite with a harmonium playing a repeating melody. Above them a ceiling fan turns at the second of its five speeds, as it has done every evening for the eleven months Deepa has been coming here. Deepa and Guruji are invented. The brain processes this paper describes in them are not.
They are practising tīntāl, a cycle of sixteen beats. Beat one, called sam, carries the strongest stress, and the cycle resolves on it. Beat nine, called khālī, is the empty beat, played without the resonant bass stroke. Deepa no longer counts on her fingers. She knows where sam is the way she knows where the next stair is in the dark. When Guruji, as a test, drops the bass stroke on a beat where she expects it, she looks up and grins. She noticed.
She does not notice the fan, whose hum is steady enough for her brain to predict it perfectly. Yet the prediction is doing work. Last month the power cut out in the middle of a lesson and the fan wound down. Deepa says the silence was “loud”. She heard the absence of a sound she had not been hearing.
So Deepa’s brain was predicting two things all evening: the fan and the tāla. Both predictions were accurate, and both were used to steer what she did next. One was part of her experience and the other was not, until it failed. What extra mechanism makes one prediction consciously experienced and leaves another merely causally useful? That is the question of this paper.
Predictive processing is the family of theories that says perception is the brain’s best guess about the causes of its sensory input, kept in check by the errors between what it predicts and what arrives. It has become one of the most widely used frameworks in cognitive science. My answer, which the rest of the paper earns, is that the framework on its own has no account of the difference. Prediction is necessary for the content of experience, and nowhere near sufficient for experience itself. The best evidence points to a conjunction of three extras: errors reaching slow, high levels of the model; high precision on those errors; and a model of one’s own perceiving that represents the content as present.
That answer has a cost, which I will state now. Every piece of evidence for it was gathered by asking people what they noticed. The fan is the case where asking fails, because the answer arrives only after the sound has gone. Whether Deepa experienced the hum all along, faintly, is exactly what report-based science cannot settle. I will come back to the fan in every part.
Guesses going down, errors going up
The modern form of the idea starts with a 1999 paper by Rajesh Rao and Dana Ballard. They built a model of visual cortex in which feedback connections carry predictions of lower-level activity and feedforward connections carry the residual errors. Trained on natural images, the model’s units developed receptive fields like those of real neurons. Some of its error units also showed effects, such as “end-stopping”, that had puzzled physiologists for decades. A neuron that fires less when a line runs past the end of its receptive field makes sense if that neuron is reporting surprise, and a long line is less surprising than a short one.
Karl Friston generalised this into the free-energy principle, which he summarised in a 2010 review in Nature Reviews Neuroscience. Andy Clark brought it to philosophers and psychologists in a long 2013 target article, “Whatever next? Predictive brains, situated agents, and the future of cognitive science”. Jakob Hohwy, Anil Seth and others extended it to attention, emotion and the self. The shared core is simple enough to state in three sentences.
First, the brain carries a hierarchical generative model: a set of guesses about what is out there and what sensory input it would produce. Second, each level sends its guess down to the level below, and each level sends up only what the guess failed to explain, the prediction error. Third, the system adjusts its guesses, or acts on the world, to reduce error over time. Perception is the guess that currently explains the input best.
A fourth idea does most of the work in explaining attention. Errors are weighted by their expected reliability, called precision. A blurry glimpse in fog produces errors the brain should mostly ignore. A clear signal in good light produces errors it should take seriously. Hohwy argued in 2012 that “conscious perception can be seen as the upshot of prediction error minimization and attention as the optimization of precision expectations”. On this view attention is not a spotlight. It is the brain turning up the gain on errors it expects to be informative.
Now return to Deepa. Her auditory cortex carries a fast, low-level model of the fan: a steady broadband sound with a slight beat at the rotation rate. That model predicts the input almost perfectly, so its errors are close to zero. Higher up, slower models carry the tāla: a sixteen-beat cycle, a stress on sam, an empty khālī. Those models predict which stroke comes next and when. When Guruji drops a bass stroke, the error at that level is large and, because Deepa is attending to the rhythm, highly weighted.
The puzzle is now sharper. On the framework’s own terms, both models are doing exactly what perception is supposed to do. Both are guesses about causes, checked against input. If perception is the best current guess, the fan is perceived. Yet Deepa says she does not hear it, and she is right in at least the ordinary sense that she could not tell you whether it was on speed two or speed three.
Analogy · the spell-checker
A spell-checker runs over every word you type and underlines only the ones it cannot match. You never see the thousands of correct guesses, only the red lines. Prediction error is like the red underline: what reaches you is the failure, not the success. Where it breaks: a spell-checker has no guesses of its own to send down. In the brain, the predictions are thought to carry the content of perception. You see the guess, not just the underlines, which is why the analogy explains the fan’s silence and fails to explain the tāla’s presence.
There are two readings of the framework that give opposite answers here, and the literature uses both without always saying which. On the first, conscious content is the winning hypothesis: whatever the brain currently takes to be out there is what you experience. Hohwy and Seth argued in 2020 that predictive processing offers a systematic basis for identifying the neural correlates of consciousness, and much work in that spirit treats the content of experience as the brain’s best current hypothesis. On the second, conscious content is the error: you experience what surprised you, and the predicted remainder stays dark. Neither reading is the framework’s official position, because the framework does not have one.
Notice what follows. If the first reading is right, Deepa hears the fan all evening and simply does not attend to it. If the second is right, she does not hear it until it stops. A theory that is compatible with both answers to the simplest question about the fan has not yet said what makes a prediction experienced. That gap is what Tomáš Marvan and Marek Havlík pressed in 2021 in “Is predictive processing a theory of perceptual consciousness?”. Their answer was that it is not, or not yet. I agree, and I want to show what the missing piece looks like.
So the question can be put precisely. Given a hierarchy of predictions and precision-weighted errors, which property separates the ones that make up experience from the ones that only steer behaviour? The literature offers four candidates, which part three examines. Before that, part two looks at the evidence that the separation is real, and that it is not where a naive reading of the framework would put it.
Local errors without awareness
The cleanest test of where prediction and awareness separate comes from a deceptively simple listening task. In 2009 Tristan Bekinschtein, Stanislas Dehaene and colleagues published “Neural signature of the conscious processing of auditory regularities”. It introduced what is now called the local-global paradigm, and it is worth describing in detail because the rest of this paper leans on it.
Each trial was five short tones, about 650 milliseconds in all. The first four were always identical to one another. The fifth was either the same, giving a pattern like AAAAA, or different, giving AAAAB. That is the local rule: does the last tone match the ones just before it? Across a block of trials, one pattern was frequent, at 80 per cent, and the other rare, at 20 per cent. That is the global rule: is this whole sequence the usual one? The trick is that the two rules can be crossed. In a block where AAAAB is the usual pattern, an AAAAA trial is locally standard, because the last tone matches, and globally deviant, because the whole sequence is rare.
The brain responds to both kinds of violation, but differently. A local violation produced an early mismatch response over auditory cortex, peaking about 130 milliseconds after the fifth tone. A global violation produced a later, widely distributed response that began at about 260 milliseconds and persisted to the end of the analysis window at 700 milliseconds. The authors’ summary is precise. Local violations led to an early response “independent of attention or the presence of a concurrent visual task”. Global violations led to a late response “only present when subjects were attentive and aware of the violations”.
The attention manipulations make the point vivid. When eleven volunteers were told to ignore the sounds and let their minds wander, the local effect was almost unchanged. The global effect was much reduced, with nothing significant beyond 400 milliseconds, and only six of the eleven showed even an early global effect individually. When volunteers did a demanding visual task instead, the global effect was not measurable at the group level. Their awareness of the global rule, scored out of four, fell from four in the attentive group to 0.4.
True story · eight patients who could not speak
The same paper tested eight people with disorders of consciousness who could not communicate. Four were in a vegetative state and four in a minimally conscious state. Three of the four vegetative patients showed the local effect, and none showed the global effect. All four minimally conscious patients showed the local effect, and three of the four showed a clear global effect. Those three later progressed to a fully conscious state over the following weeks. The authors were careful in the other direction too: the global effect “can be absent in conscious subjects who are not aware of the global auditory regularities”. A missing global response is not proof that no one is there.
Sleep tells the same story from another angle. In 2015 Melanie Strauss and colleagues recorded the local-global paradigm across a night. In “Disruption of hierarchical predictive coding during sleep” they reported that “the global response, reflected in the P300, vanished during sleep”. The local mismatch response remained in light sleep, deeper sleep and dream sleep, “but with an incomplete structure”: a specific peak reflecting prediction error disappeared. Their conclusion was that sleep leaves initial auditory processing intact “but specifically disrupts both short-term and long-term auditory predictive coding”.
Put the two studies together. Low-level prediction error survives inattention, sleep and the vegetative state. High-level prediction error, the kind that tracks a regularity spread over several seconds, survives none of them reliably. Whatever makes a prediction conscious, it is not the mere presence of a prediction error somewhere in the brain. If it were, a vegetative patient with a clean mismatch response would count as aware of the deviant tone, and no one, including the authors, draws that conclusion.
This also answers part of Deepa’s puzzle, because the fan is a local regularity. Its model sits low in the auditory hierarchy, near the level where mismatch responses live. The tāla is a global regularity, a pattern over sixteen beats, and its model sits higher. On the local-global evidence, it is the depth of the model that goes with awareness. The fan is predicted at a depth that never reaches experience while it works.
A hierarchy of predictions, not just of errors
A sceptic could say the mismatch response is not prediction error at all, only neurons tiring of a repeated sound. Wacongne and colleagues designed a test for exactly that in 2011. Their participants heard frequent series of four identical tones followed by a different fifth tone. Because the change itself was expected, a hierarchical predictive model says its mismatch response should be partly cancelled by a higher prediction. It also predicts that omitting the fifth tone should produce a larger error when a different tone was expected. Their EEG and MEG recordings confirmed these predictions, and the authors concluded that “higher-order predictions appear to be generated in multiple areas of frontal and associative cortices”. Simple adaptation cannot explain that pattern, so the brain really does predict at several levels at once.
Expectation changes what reaches awareness
The second body of evidence asks whether prediction affects entry into awareness at all. It does, and in a particular direction. Lucia Melloni and colleagues showed in 2011 that expectations lower the threshold of conscious perception and shift the brain’s seen-versus-unseen difference earlier: about 300 milliseconds after a stimulus without expectations, about 200 milliseconds with them. Yair Pinto, Simon van Gaal, Floris de Lange, Victor Lamme and Anil Seth used continuous flash suppression. In this method an image shown to one eye is hidden by flashing patterns in the other. They found that expected stimuli enter consciousness faster than neutral or unexpected ones. Timo Stein and Marius Peelen found that a valid category cue, such as “car”, improves detection by raising perceptual sensitivity, not just by shifting willingness to respond.
Pinto and colleagues drew the moral explicitly: their findings “support the hypothesis that conscious access depends on verification of perceptual predictions”. That favours the first reading from part one, on which content is the confirmed hypothesis rather than the error. It also sits awkwardly with the fan, which is the most verified prediction in Deepa’s room.
When the prior wins
The last result in this part comes from psychiatry and is the most uncomfortable. Albert Powers, Christoph Mathys and Philip Corlett published “Pavlovian conditioning-induced hallucinations result from overweighting of perceptual priors” in 2017.
True story · hearing a tone that was not played
Powers and colleagues paired a visual stimulus with a sound until the two were associated. Later, when the visual stimulus appeared alone, some people reported hearing the sound. They tested four groups of people who differed in whether they heard voices in daily life and whether they sought treatment for it. “People who hear voices were significantly more susceptible to the effect.” With brain imaging and a computational model of perception, the team traced the effect to a circuit and to how strongly people weighted their prior expectations. A strong enough prediction, with no matching input, was experienced as a sound.
A 2019 review by Corlett and colleagues, “Hallucinations and strong priors”, describes the wider pattern: “strong, overly precise priors can engender hallucinations in healthy subjects”. This is the mirror image of the fan. The fan is a well-supported prediction that is not experienced. The conditioned tone is an unsupported prediction that is. Whatever the extra ingredient is, it cannot be the match between prediction and input, because the conditioned tone has no input at all. It looks much more like the precision the system assigns to its own prediction, at a level that reaches awareness.
So part two leaves three constraints, and the first is that low-level errors are not enough. Expectations speed entry into awareness when they are confirmed. And a sufficiently precise high-level prediction can be experienced with no input. Any candidate for the extra ingredient has to satisfy all three, and has to say something about the fan.
Four candidates for the extra ingredient
If prediction alone does not make experience, something must be added. Four proposals recur in the literature, and each has serious defenders. Each picks out a real property of predictive systems, and each has a case it handles badly. I take them in turn and then say which combination I think survives.
Candidate A: precision
The first proposal is that a prediction or error becomes conscious when it is weighted with high enough precision. This is the natural reading of Hohwy’s view that attention is precision optimisation. Errors you expect to be reliable get amplified, win the competition with rival hypotheses, and settle into what you experience. Errors you expect to be noise get damped and stay in the dark.
Precision handles the conditioned hallucinations well. An overly precise prior can override weak input, or no input, and be experienced. It also explains some attention effects in the local-global data, since the global response weakens when attention is drawn away and precision on the auditory stream drops.
It handles the fan badly, because the fan’s model is among the most precise in the room. It has been confirmed every evening for eleven months, and its errors are reliably near zero. If high precision made a prediction conscious, the fan would be vivid. One reply is that precision attaches to errors, not predictions, and the fan produces no error to weight. That reply moves the problem rather than solving it, because it turns precision into a criterion for surprise and leaves us with the second reading from part one.
There is a neural complication too. Peter Kok, Janneke Jehee and Floris de Lange showed in 2012 that expectation reduces the response in primary visual cortex while improving the stimulus representation there, as measured by pattern decoding. The authors’ title says it in three words: “Less is more”. An expected input is represented more sharply with less activity. So the brain’s best-predicted contents may be its quietest, and any criterion based on signal strength will miss them.
What this is for: seeing how precision decides whether the expectation or the input wins. What it is not: a model of experience; the percept curve is a posterior, and whether it is felt is the question.
With JavaScript off: at low input precision the percept stays near the expectation; at high precision it moves to the input. Equal precisions put it halfway.
Analogy · the mixing desk
A sound engineer at a concert does not change what the musicians play. She moves faders, so that some channels reach the hall loudly and others barely at all. Precision works like a fader on each stream of error. Where it breaks: a fader that is turned down still passes some sound, and nobody in the hall decides which channel they hear. The analogy explains how attention can change awareness without changing the model, but it has no listener, which is the very thing in question.
Candidate B: hierarchical depth and broadcast
The second proposal is that a prediction becomes conscious when it reaches high enough in the hierarchy, where slow models integrate information across senses and time. This is what the local-global data most directly support. It also connects predictive processing to global workspace theory, which says a content becomes conscious when it is broadcast across a network of long-range neurons.
Christopher Whyte and Ryan Smith built the connection formally. In “The predictive global neuronal workspace”, published in 2021, they cast workspace dynamics as approximate Bayesian inference in an active inference model. They report that it reproduces findings on inattentional blindness and the four-way taxonomy of conscious, preconscious, subliminal and inattentive processing that workspace theory predicts. In their model, a content becomes conscious when it is inferred at a deep level that integrates over time and feeds into action and report.
Depth handles the fan well, and handles Deepa’s tāla well. It handles the patients well: local responses without global ones go with absence of awareness. Its weakness is that it is hard to state depth without stating access. If the deep level is defined as the level whose contents can be reported, the criterion becomes “conscious predictions are the reportable ones”. That is a description of the data, not an explanation, and it inherits every difficulty of workspace theory with experience that might be present but not reported.
Candidate C: temporal thickness and counterfactual depth
The third proposal comes from Friston himself and from Seth. In a 2018 essay, “Am I self-conscious?”, Friston proposed that “the mind comes into being when self-evidencing has a temporal thickness or counterfactual depth, which grounds inferences about the consequences of my action”. Seth had argued in 2014 that the sense that perceived objects are really there, which he called perceptual presence, depends on “counterfactually-rich” generative models. These are models that encode how input would change under actions the system could take, even if it does not take them. He used this to explain why synaesthetic colours are vivid yet not experienced as being on the page.
Temporal thickness handles the tāla very well. Deepa’s model of the cycle extends over many seconds and supports counterfactuals: if Guruji drops a stroke, she knows where she is; if she speeds up, she knows what comes next. The fan’s model is thin. It predicts the next moment from the last and supports almost no counterfactuals, since Deepa does nothing that changes the hum.
Its weakness is that it explains a dimension of experience, its depth and presence, better than the boundary between experience and its absence. A thin model might still give a thin experience. Deepa, asked to attend to the fan, will say she hears it at once, although its model has not suddenly become counterfactually rich. Temporal thickness looks like a grade, not a gate.
Candidate D: inference about one’s own perceiving
The fourth proposal treats awareness as a second inference, about the first. Stephen Fleming put it this way in 2020 in “Awareness as inference in a higher-order state space”: reports of awareness are “metacognitive decisions (inference) about a generative model of perceptual content”. The states that support those decisions are simple. They vary along a single line from absent to present. Fleming also argued that the architecture is asymmetric, with many perceptual states nested under “present” and few under “absent”, and showed in simulations that this asymmetry yields something like the “global ignition” seen in brain imaging of awareness reports.
Inference about one’s own perceiving handles the fan in an interesting way. The fan’s content is represented, accurately, at the first level. At the second level, the system has no reason to represent “I am hearing a fan” as present, because nothing calls for that inference. When Deepa attends to the hum, or when it stops, the second-level inference flips, and she reports it. On this view the hum was perceived and not experienced, or perceived and not represented as perceived, which may come to the same thing.
Its weakness is the one every higher-order view faces. It seems to deny experience to creatures and states that lack the second-order model, such as infants, many animals and, perhaps, absorbed adults. And it makes experience depend on a representation that could, in principle, misfire: a system could represent itself as seeing something it does not represent at the first level.
What this is for: comparing how each candidate criterion sorts the paper’s cases. What it is not: a measurement; the verdicts are my reading of each candidate.
With JavaScript off, all seven notes:
- Fan hum: Not reported until it stops. Only temporal thickness gives it anything, and only a little.
- Tāla cycle: Reported and attended. All four candidates agree, which is why it teaches nothing.
- Sleeper, local deviant: Local errors survive sleep. Only precision, read loosely, gives a partial yes.
- Vegetative patient with a global effect: Depth says yes, higher-order inference cannot say without report. The uncomfortable case.
- Conditioned hallucination: A tone is heard that was not played. Every candidate says the percept is conscious, which fits the reports.
- Bayesian air conditioner: The negative control. A criterion that grants it awareness is too cheap.
- Nadi earbuds: One fast level, fixed gain, no self-model. Predicts like the fan model, not like the tāla model.
What survives
None of the four works alone. Precision fails on the fan, and depth is close to a restatement of report. Temporal thickness is a grade, not a gate. Higher-order inference handles the fan but threatens to deny experience to the absorbed and the young.
A conjunction does better. A prediction is a candidate for experience when its errors reach a temporally deep level of the model, are weighted with high precision there, and are represented by a further model as present to the system. Depth and thickness pick out where the content lives, and precision decides whether it wins. The higher-order model decides whether the system treats it as its own current perceiving. Deepa’s tāla meets all three. The fan meets the first only weakly and the other two not at all, until it stops, when all three switch on together.
I want to be clear about what this conjunction is. It is the best summary of the evidence I can construct, and it fits the local-global data, the hallucination data and the fan. It is not an explanation of why such a conjunction should be experienced at all. And every one of its three clauses was calibrated against what people report. That last point becomes the main objection in the strongest case against, and I do not think it can be fully answered.
What other theories make of the fan
Predictive processing is not, strictly, a theory of consciousness. It is a theory of how brains infer, which several theories of consciousness have borrowed. That makes the rival positions unusual, since each can absorb the framework on its own terms. Most of them can be restated in predictive language, and the question is whether the restatement adds anything. A good test is to ask each what it says about Deepa’s fan.
Global neuronal workspace theory says a content becomes conscious when it is broadcast by a network of long-range neurons concentrated in prefrontal and parietal cortex. Its authors would read the local-global data as direct support. Bekinschtein, Dehaene and colleagues designed the paradigm partly to test the idea that conscious processing requires sustained, widely distributed activity, and the global effect has that profile. On this theory the fan was processed and never broadcast, so it was not experienced. When it stopped, the error was large enough to ignite the workspace. The predictive global workspace of Whyte and Smith keeps all of this and adds a formal account of how expectations shift the threshold.
Higher-order theories say a state is conscious when it is represented by a suitable higher-order state, roughly a representation of oneself as being in it. Hakwan Lau and David Rosenthal reviewed the case in 2011 in “Empirical support for higher-order theories of conscious awareness”. Fleming’s model is a predictive version of this. The fan was perceived without a higher-order representation of perceiving it, so it was not conscious. The higher-order theorist welcomes candidate D and treats the other three as describing what is being represented, not what makes the representing conscious.
Recurrent processing theory, the subject of the previous paper in this series, says local feedback within sensory cortex is enough for experience. It is the rival that would most readily say Deepa heard the fan all evening. Auditory cortex was certainly running a feedback loop that modelled the hum. On Lamme’s view, that could be phenomenal experience without access. Notice that Lamme is a co-author of the Pinto study, which found expectations speed entry into awareness. Recurrent processing theory can accept that prediction shapes what becomes conscious while denying that report marks the boundary.
Integrated information theory is the outlier. It holds that experience corresponds to the intrinsic cause-effect structure of a physical system, not to any computation it performs. On this view, whether a prediction is conscious depends on the integrated structure of the neurons carrying it at that moment, and the predictive role of those neurons is irrelevant. The fan would be conscious if the substrate carrying its model were part of the brain’s maximally integrated complex at the time, and not otherwise. That is a clear answer, though not one anyone can currently check.
The deflationary rival
The last rival does not offer a different extra ingredient, because it denies there is one to find. Keith Frankish’s 2016 target article, “Illusionism as a theory of consciousness” in the Journal of Consciousness Studies, argued that phenomenal consciousness, as philosophers usually conceive it, is an introspective illusion. On this view the task is to explain why some states seem to have a special inner quality, not to find the quality.
Illusionism fits predictive processing well, and some predictive theorists lean towards it. If experience is the brain’s model of its own states, and that model can be wrong, then the conscious predictions are simply those the system models itself as having. There is no further fact about whether Deepa heard the fan. There is only whether her self-model represented a hearing. The question this paper asks becomes a question about the self-model, which is candidate D with the metaphysics removed.
Critics of the framework itself
Some critics doubt the framework more than any theory of consciousness built on it. Kevin Walsh, David McGovern, Andy Clark and Redmond O’Connell reviewed the neurophysiological evidence for predictive processing as a model of perception in 2020. They noted that the models have been “criticized for lacking the empirical support to justify their status”, partly because their unique predictions are hard to test. Paweł Litwin and Marcin Miłkowski were blunter in “Unification by fiat: arrested development of predictive processing”, arguing that the framework’s claim to unify the mind has been asserted rather than earned.
These critiques matter for this paper in a specific way. If predictive processing can redescribe any finding, then the four candidates in part three are not discoveries about prediction. They are older ideas about attention, broadcast, temporal integration and self-representation, renamed. I think that is partly true, and I return to it in the strongest case against. The part that is not true is the local-global result. That paradigm was designed around a hierarchical prediction, made a risky forecast, and the forecast came true.
So the rivals sort cleanly on the fan. Workspace theory and higher-order theory say it was not experienced while it worked. Recurrent processing theory says it may have been. Integrated information theory says it depends on structure, not on prediction. Illusionism says there is no fact to find beyond the self-model. Predictive processing, as a framework, can host all five, which is a measure of its flexibility and of how much work the extra ingredient has to do.
The carpenter’s mark and the unrecognised lover
Two Indian traditions had detailed theories of perception that put something like prediction at the centre. Neither knew anything about cortex or neurons. Both had a word for the act of matching what is present with what was known before, and both thought that act was where experience gets its shape. Reading them against predictive processing is useful because they disagree about whether matching is enough.
Saññā in the Abhidhamma
In the Theravāda Abhidhamma, every moment of consciousness, called a citta (चित्त), arises with a set of mental factors. One of them, present in every citta without exception, is saññā (सञ्ञा), usually translated as perception or recognition. Buddhaghosa’s fifth-century Visuddhimagga defines it in a passage, XIV.130, that reads almost like a description of a prior. In Bhikkhu Ñāṇamoli’s translation, its function is “to make a sign as a condition for perceiving again that ‘this is the same’, as carpenters, etc., do in the case of timber”. I take the wording from a transcription of the passage that gives the Pāli alongside Ñāṇamoli’s English, which should be checked against the printed edition.
The passage continues with two more images. Saññā is “manifested as the action of interpreting by means of the sign as apprehended, like the blind who ‘see’ an elephant”. And its proximate cause is the object as it appears, “like the perception that arises in fawns that see scarecrows as men”. The fawns are the striking case. A scarecrow is taken as a man because the sign of “man” has been laid down before and the present shape fits it. That is perception as the application of a learned model, with all the confidence and all the error that implies.
Here the parallel with predictive processing holds well. Saññā is a learned template, laid down by past experience, applied to present input, and capable of producing confident mistakes. The fawns and the scarecrow are the Abhidhamma’s conditioned hallucination.
And here it breaks. Saññā arises with every citta, and every citta is, by definition, a moment of consciousness. The Abhidhamma has no category of unconscious perception in the modern sense. Even the bhavaṅga (भवङ्ग), the life-continuum that flows between active cognitive processes, is a series of cittas with its own object. So the tradition cannot pose this paper’s question in its own terms. For it, the question is not which predictions are conscious but which objects initiate a full cognitive process and which do not.
That reframing is still useful. In the eye-door process described in the previous paper, a very slight object only makes the life-continuum vibrate, and no cognitive process follows. On my reading, though no commentary I know says this, Deepa’s fan is closer to that case than to unconscious perception. It is an object that impinges and does not start a process. When it stops, the change is a new object, strong enough to start one. That reading is mine and should be treated as a proposal, not as the tradition’s view.
Recognition in Pratyabhijñā
The Pratyabhijñā school of Kashmir Śaivism takes its name from the word pratyabhijñā (प्रत्यभिज्ञा), recognition: knowing again, “this is that”. Its central text is Utpaladeva’s tenth-century Īśvarapratyabhijñākārikā, the Verses on the Recognition of the Lord. Near its end, Utpaladeva gives an example that his own commentary uses to establish the whole doctrine. In the GRETIL e-text with his own commentary it is verse 4.17; other editions number it 4.1.17. Raffaele Torella’s critical edition and translation remains the scholarly standard.
Primary text · Īśvarapratyabhijñākārikā 4.17
तैस्तैरप्युपयाचितैरुपनतस्तन्व्याः स्थितोऽप्यन्तिके
कान्तो लोकसमान एवमपरिज्ञातो न रन्तुं यथा ।
लोकस्यैष तथानवेक्षितगुणः स्वात्मापि विश्वेश्वरो
नैवालं निजवैभवाय तदियं तत्प्रत्यभिज्ञोदिता ॥
taistair apy upayācitair upanatas tanvyāḥ sthito ’py antike / kānto lokasamāna evam aparijñāto na rantuṃ yathā / lokasyaiṣa tathānavekṣitaguṇaḥ svātmāpi viśveśvaro / naivālaṃ nijavaibhavāya tad iyaṃ tatpratyabhijñoditā
My translation: “A lover, brought near by her many entreaties, may stand right beside the slender woman. Yet while he is unrecognised, he seems like any other man and cannot delight her. In the same way, the Lord of all, though he is people’s own self, cannot show his own splendour while his qualities go unnoticed. That is why this recognition has been taught.”
Utpaladeva’s own short commentary adds that “by the example of the lover” the recognition of the Lord is to be established, and that its fruit is “unprecedented” (apūrvaphalā). The lover is fully present, seen and heard in the room. Everything a perceptual model could supply about him is supplied. What is missing is the act of recognition that connects the present man with what she already knows of him. Until that act, his presence does nothing for her.
Where the parallels hold and where they break
Both traditions hold the same structure: perception involves applying something learned to something present, and the act of matching can go wrong. That fits predictive processing well enough. Both also locate something important beyond mere modelling. In the Abhidhamma, it is whether an object starts a full process with impulsion (javana). In Pratyabhijñā, it is recognition, and behind recognition the reflexive awareness (vimarśa, विमर्श) that Utpaladeva says is the very nature of light.
The breaks are just as clear. The Abhidhamma has no unconscious perception, so it cannot host the paper’s question without reframing it. Pratyabhijñā is about recognising the self as Śiva, and its metaphysics makes consciousness prior to any model. It also matters that, for Utpaladeva, vimarśa is not a second-order representation. It is the svabhāva, the own nature, of light itself. That puts the tradition at odds with Fleming’s higher-order model, even though both say awareness needs more than first-order content.
What this means for RTM
RTM, the Recognition Theory of Mind this series tests, uses Utpaladeva’s terms for its layers. Its L1 is spanda, responsiveness; L2 prakāśa, manifestation; L3 vimarśa, reflexivity; L4 pratyabhijñā, recognition; L5 cit, the luminous ground. Predictive processing gives a useful way to test it, because the four candidates in part three map onto RTM’s layers in a definite way.
Precision and depth, candidates A and B, describe which content shows up and how widely it is available. That is L2. Inference about one’s own perceiving, candidate D, is the closest scientific analogue to L3. Temporal thickness, candidate C, is closest to L4, because recognition in Utpaladeva’s sense joins present input to a model built over time: “this is that”. If the mapping is right, RTM makes a prediction that its rivals do not. Experience of a content should require L3 and should be deepened, not created, by L4. Deepa’s fan, on this reading, has L2 without L3 while it works, and gains L3 when it stops.
The tradition also forces a correction on RTM. Suppose vimarśa is the nature of light rather than a second representation. Then RTM’s L3 should not be modelled as a higher-order layer that could be switched off while L2 runs on, as in the device thought experiment in part four. RTM should say L3 is same-order reflexivity, built into manifestation, and should predict that clean dissociations between first-order content and awareness are rare and fragile. That is a risky prediction. Blindsight, the subject of a later paper in this series, is its first hard test.
A toy brain on the local-global test
Arguments about criteria are cheap until a criterion has to sort real cases. So I built the smallest predictive system I could that still has two levels, ran it on the local-global paradigm, and asked two candidate criteria to say which conditions involve awareness of a violation. The point is not to model a brain. It is to see whether a criterion that sounds plausible survives a negative control, and where it breaks.
The model has two levels. Level 1 predicts that each tone will match the one before it. Its error at the fifth tone is 1 when the tone changes and 0 when it does not. Level 2 learns, across trials, how often each five-tone pattern occurs. Its error on a trial is the surprise of that pattern, measured in bits, multiplied by a precision weight that stands for attention. Inattention lowers that weight and slows learning, while sleep is modelled crudely by switching level 2 off. A negative control runs the same model on sequences where both patterns are equally common, so there is no global rule to violate.
Two criteria are tested. Criterion C1 says a condition involves awareness of a violation if any prediction error is present at any level. Criterion C2 says it does only if the precision-weighted level-2 error exceeds a threshold, set at 1.5 bits, on more than 5 per cent of trials. The truth column is taken from the human data in part two: attentive listeners are aware of global violations, mind-wanderers partly, sleepers not, and in the negative control there is nothing to be aware of.
The model’s code, lightly condensed for display with the same logic, and its actual output
def run(condition, n_trials=400, p_rare=0.2, attention=1.0,
level2_on=True, regular=True):
counts = {'xxxxx': 1.0, 'xxxxY': 1.0} # prior over patterns
frequent = 'xxxxY' # AAAAB is the usual pattern
rows = []
for t in range(n_trials):
pat = (frequent if random.random() > p_rare else 'xxxxx') if regular \
else random.choice(['xxxxx', 'xxxxY'])
local_pe = 1.0 if pat.endswith('Y') else 0.0
if level2_on:
p = counts[pat] / sum(counts.values())
global_pe = -math.log2(p) * attention
counts[pat] += attention
else:
global_pe = 0.0
rows.append((pat, local_pe, global_pe))
return rows[50:] # drop learning phase
# output (random seed 36)
attentive localPE freq=1.00 rare=0.00 | globalPE freq=0.30 rare=2.46 | flagged=0.19
mind-wandering localPE freq=1.00 rare=0.00 | globalPE freq=0.10 rare=0.72 | flagged=0.00
sleep localPE freq=1.00 rare=0.00 | globalPE freq=0.00 rare=0.00 | flagged=0.00
negative control localPE freq=1.00 rare=0.00 | globalPE freq=1.26 rare=0.79 | flagged=0.00
criterion test: does each candidate criterion mark the condition 'aware of a violation'?
attentive truth=yes C1 any error=yes C2 deep weighted error=yes
mind-wandering truth=partly C1 any error=yes C2 deep weighted error=no
sleep truth=no C1 any error=yes C2 deep weighted error=no
negative control truth=no (nothing to detect) C1 any error=yes C2 deep weighted error=no
fan hum truth=disputed C1 any error=no C2 deep weighted error=noIn the negative control the labels “freq” and “rare” are only names, since both patterns occur half the time. The fan hum row is a constant input with no change, entered by hand: every error is zero by construction.
The output makes three points. First, C1 fails the negative control and the sleep condition. Local error is present in every condition, because the fifth tone changes on most trials whatever the listener is doing. A criterion based on the presence of prediction error would say sleepers, and the toy model in a meaningless sequence, are aware of violations. That is the toy version of the vegetative patients with intact mismatch responses.
Second, C2 gets attentive, sleep and the control right. It flags about 19 per cent of trials in the attentive condition, close to the 20 per cent of trials that really were globally deviant, and none in sleep or the control. This is the positive case: the instrument detects what it should when it should.
Third, C2 fails on mind-wandering. It says “no”, but the human data say “partly”: six of Bekinschtein’s eleven mind-wanderers showed an early global effect. In the model, precision-weighted surprise on rare trials falls to 0.72 bits, below the threshold. The criterion turns a graded fact into a cut, and the cut falls in the wrong place for half the people. That is a counterexample to C2 as a gate.
What this is for: finding where criterion C2 stops working as its threshold moves. What it is not: a principled threshold; the window was found by checking against the answer.
With JavaScript off: the chart shows every threshold. The criterion works only between about 1.5 and 2 bits.
The threshold itself is the failure boundary. I ran the same four conditions with thresholds from 0.5 to 3 bits. Below about 1.5 bits, C2 starts flagging the negative control: at 1 bit it marks 46 per cent of control trials, and at 0.5 bits all of them. Above about 2 bits it starts missing the attentive listener: at 2.5 bits it catches only 5 per cent. The criterion works in a window from roughly 1.5 to 2 bits, and nothing in predictive processing says where that window should be. It was chosen by checking against the answer.
What this is for: watching the deep error grow with attention while the local error stays flat. What it is not: data from people; attention here is one multiplier in a toy model.
With JavaScript off: the rare-pattern error crosses the 1.5-bit line between attention 0.6 and 0.7.
Worked case 1: Deepa’s dropped stroke
Apply the three questions from the TLDR to the moment Guruji drops a bass stroke. At what depth is the error? The missing stroke violates a sixteen-beat model that Deepa holds over several seconds, which is a deep, slow level. With what precision? Deepa is attending to the rhythm, and the error falls on a beat whose identity she is certain of, so precision is high. Does a further model represent the content as present? She looks up and grins, which is a report, and reports are what such a model is for. All three conditions hold. The conjunction from part three says the violation is experienced, and Deepa agrees.
Now apply the same questions to the fan before it stops. The error is near zero at a low level, precision on it is irrelevant because there is nothing to weight, and no further model represents the hum as present. The conjunction says it is not experienced, which is the expected answer. Whether it is the right answer is the open question this paper cannot close.
Worked case 2: the earbuds that “hear like you do”
An invented audio start-up in Hyderabad, Nadi Audio, sells earbuds with adaptive noise cancelling. Their marketing says the chip “predicts the world around you, just as your brain does, and hears like you do”. A journalist asks whether that means the earbuds are aware of sound. Here is how the instrument and the conjunction apply.
The chip runs a predictive filter. It models incoming noise over a few hundred milliseconds and cancels the predicted part. That is level 1 in the toy model: a fast, local predictor. Its errors drive the update of the cancelling filter. It has no slow model of sequences, no precision that varies with any analogue of attention beyond a fixed gain schedule, and no model of its own predicting.
- Depth: shallow, one level with a short window. Fails.
- Precision: fixed weights, no analogue of attention. Fails.
- Self-model: none. Fails.
- C1, any prediction error: passes, which is exactly why C1 is useless.
The honest answer to the journalist is that the earbuds predict like the fan model in Deepa’s auditory cortex, not like her model of the tāla. “Predicts like your brain” is true of the part of the brain that predicts in sleep. That does not make the marketing claim false. It makes it uninformative about awareness, which is the decision this paper recommends: treat prediction alone as causally useful, not as evidence of experience.
The case that prediction is beside the point
Here is the objection as its best defenders would make it. It comes in three steps, and I think the first two are right.
Step one: predictive processing contributes nothing to the answer. Look at the conjunction this paper defends, clause by clause. Depth and broadcast come from workspace theory, and precision is attention under another name. Temporal integration is an old idea from the psychology of time. The model of one’s own perceiving is higher-order theory. Every ingredient that does the work was available before anyone spoke of prediction error. Saying that conscious predictions are deep, precise and self-modelled is saying that conscious contents are broadcast, attended and represented as one’s own, with “prediction” added in front. This is the force of the “unification by fiat” complaint, and it lands.
There is an empirical version of the same worry. Kevin Walsh, David McGovern, Andy Clark and Redmond O’Connell reviewed the neurophysiology in 2020 and opened by noting that predictive processing models are “criticized for lacking the empirical support to justify their status”. They traced part of the gap to how hard it is to test the predictions unique to these models. Their review then assessed a newer wave of studies designed to fill it. The point for this paper is narrower than their review. Most findings that fit prediction also fit older accounts. A mismatch response can be read as adaptation to the repeated tone, and a late global response as attention catching a rare event. The local-global result survives this worry better than most, because the rare AAAAA trial is locally the most adapted sound in the block.
Step two: the criterion is circular, because each of its clauses was fixed by checking against report. The local-global global effect was defined by contrast with what aware listeners say. The threshold in the toy model was chosen to sort conditions whose truth was taken from report. Fleming’s higher-order model is explicitly a model of awareness reports. So the conjunction is a model of reportable inference. That is a fine thing to have, but it cannot tell us whether the fan was experienced, because the fan is the case where report is silent. The theory cannot lose on the cases that matter most.
Step three: the question is malformed. On a deflationary view, there is no fact about whether a prediction is “consciously experienced” over and above whether the system models it as such. Then this paper’s question dissolves into the question of which predictions the self-model tracks, and the answer is whatever the self-model tracks. The fan was not experienced because the self-model did not represent it. End of story.
Analogy · the satnav that says “recalculating”
A car’s satnav predicts the road all the way, and the driver hears from it only when the car leaves the route. Its whole model of the drive sits behind one word, “recalculating”. That is the deflationist’s picture: what reaches you is what the self-model announces. Where it breaks: the satnav’s announcement changes nothing about the road, while the hallucination data suggest the brain’s self-model and its first-order model change together. The analogy fits illusionism too neatly to be evidence for it.
What survives? I concede step one in part, since the ingredients are not new. What predictive processing adds is a common currency. Depth, precision and self-modelling can all be stated as properties of one kind of inference, which makes them comparable and lets them be built into a single model and tested together, as Whyte and Smith did. The local-global paradigm is the proof that this currency can buy something. It was designed from a hierarchical predictive idea, made a risky forecast about which errors survive inattention, and was right.
I concede step two in full. The criterion is a model of reportable inference. The paper’s answer to its own question is therefore conditional: if experience goes with what can in principle be reported, then the extra mechanism is deep, precise, self-modelled inference. If there is experience that cannot be reported, this paper has nothing to say about it, and neither does any theory calibrated the same way.
I reject step three, for a reason that is empirical rather than metaphysical. If the self-model alone decided, then switching off the self-model should change nothing but report. The conditioned hallucination data suggest that it changes more than report. People with strong priors report hearing a tone, and their behaviour, their confidence and their brain activity all shift together. That is not proof against illusionism. It is evidence that the self-model and the first-order model are entangled, so the clean case the deflationist needs may not exist.
What would change my mind
My central claim is that prediction error minimisation alone does not separate conscious from unconscious inference, and that the best-supported extra is a conjunction of depth, precision and self-modelling. Four observations would move me.
A reliable global effect in deep dreamless sleep or in confirmed unconsciousness would weaken the depth clause. If slow, high-level prediction errors over a sequence turned up in people who, by every other test, were not aware, depth would be shown to be insufficient.
A clean double dissociation between first-order content and awareness reports, under conditions where the higher-order model is selectively disrupted, would favour the higher-order reading over the conjunction. It would also count against RTM’s claim that L3 is same-order reflexivity. That claim predicts such dissociations are rare and fragile.
Picture a paradigm that detects a precision-weighted deep error to a background sound, such as a fan, while the person denies hearing it. Asked to attend, the person then describes the sound in detail they could not have guessed. That result would push me towards the view that background contents are experienced and unreported. It would also show that the conjunction draws the line too high.
A predictive model with all three properties that gave clearly wrong answers on human data would count against the conjunction directly. The toy model in part six is too simple to be that test. A serious one would need a model like the predictive global workspace, run blind on a new paradigm.
Some things would not move me, and it is fair to say which. Another demonstration that expectation changes what people report would not, because every theory now predicts that. A new mathematical restatement of consciousness as free-energy minimisation would not either, unless it came with a case where it forecasts a result the rivals get wrong. And a report from contemplatives that background sound is experienced would not settle the fan on its own. Such reports are themselves reports, made in a trained state that changes attention. I would take them seriously as data about trained attention, which is a different question from what an untrained listener experienced on an ordinary evening.
What I still don’t know
I don’t know how to set the depth threshold except by checking against report, and I don’t know whether there is a principled way to do it. I don’t know whether temporal thickness is a gate or a grade, and the fan suggests a grade. I don’t know whether the higher-order model is a separate system that could fail on its own or part of first-order perception. That question decides whether RTM’s same-order L3 survives. And I don’t know whether Deepa heard the fan. I have stopped thinking that question is easy, which is something.
Claim Ledger
Each claim has an identifier for the programme’s ledger, a confidence level, the kind of evidence behind it and the observation that would refute it.
| ID | Claim | Confidence | Evidence type | What would refute it |
|---|---|---|---|---|
| C1 | The brain generates predictions and prediction errors at several hierarchical levels. | Established | Published evidence | Mismatch and omission effects fully explained by adaptation alone. |
| C2 | Low-level mismatch responses survive inattention, sleep and the vegetative state. | Established | Published evidence | Failure to replicate local effects in these states. |
| C3 | The presence of prediction error is not a marker of awareness. | Established | Published evidence, argument | Evidence that sleepers or vegetative patients are aware of local deviants. |
| C4 | Slow, global prediction errors go with awareness of the violated regularity. | Likely | Published evidence | A reliable global effect in confirmed unconsciousness. |
| C5 | Confirmed expectations speed entry into awareness. | Likely | Published evidence | Failures to replicate with response bias fully controlled. |
| C6 | Precision alone cannot be the criterion for conscious prediction. | Likely | Argument, toy model | A principled precision measure that sorts the fan and the tāla correctly. |
| C7 | The best-supported extra is a conjunction of depth, precision and self-modelling. | Speculative | Synthesis | A model with all three giving wrong answers on new human data. |
| C8 | Every current criterion is calibrated against report and cannot decide unreported cases. | Established | Argument | A validated marker of experience independent of report. |
| C9 | RTM’s L3 should be modelled as same-order reflexivity, predicting rare and fragile content-awareness dissociations. | Speculative | Textual reading, hypothesis | Reliably repeatable double dissociations under selective disruption. |
| C10 | Saññā and pratyabhijñā both place matching of past and present at the centre of perception, and both locate experience beyond matching. | Likely | Textual reading | Standard commentaries treating either as the whole of cognition. |
One dated prediction joins the programme’s Prediction Registry. By 31 December 2030, at least one peer-reviewed study will report a measurable global local-global effect in some patients classified as vegetative or unresponsive by behavioural examination, and those patients will have better outcomes on average than matched patients without it. It will be judged against studies indexed in PubMed using the local-global paradigm in disorders of consciousness. If it fails, the depth clause loses some of its clinical value.
Glossary
- Bhavaṅga (भवङ्ग)
- Pāli, “life-continuum”: in the Abhidhamma, the stream of consciousness between active cognitive processes.
- Citta (चित्त)
- A single moment of consciousness in the Abhidhamma analysis. Every citta is conscious by definition.
- Continuous flash suppression
- A method that hides an image shown to one eye by flashing changing patterns to the other eye.
- Counterfactual depth
- How far a model encodes what would happen under actions the system could take but does not.
- Generative model
- A model of what causes sensory input, which can be run to predict that input.
- Global effect
- The late, widespread brain response to a violation of a rule spread across trials in the local-global paradigm.
- Local-global paradigm
- A listening task that crosses a within-trial rule (does the last tone match?) with an across-trial rule (is this the usual pattern?).
- Mismatch response
- An early brain response, around 100 to 200 milliseconds, to a sound that breaks a short-term pattern.
- Precision
- The expected reliability of a signal or error. High-precision errors are given more weight.
- Prediction error
- The part of the input that the current prediction does not explain.
- Pratyabhijñā (प्रत्यभिज्ञा)
- Sanskrit, “recognition”, knowing again: “this is that”. The name of Utpaladeva’s school and RTM’s layer L4.
- Prakāśa (प्रकाश)
- Sanskrit, “light”: the sheer showing-up of anything in consciousness. RTM’s layer L2.
- Saññā (सञ्ञा)
- Pāli, perception or recognition: the mental factor that marks an object so it can be known again.
- Spanda (स्पन्द)
- Sanskrit, “vibration”: in Kashmir Śaivism, the dynamic pulse of consciousness. RTM’s layer L1, responsiveness.
- Temporal thickness
- How far into the past and future a model’s inferences extend.
- Vimarśa (विमर्श)
- Sanskrit, reflexive awareness: light grasping itself. RTM’s layer L3.
When the fan stops
A week after the power cut, Deepa tries an experiment of her own. In the middle of a cycle she asks Guruji to keep playing while she turns her attention to the fan. She hears it at once: a low hum with a slight beat, steadier than she expected. She loses sam for two whole cycles while she listens. When she returns to the rhythm, the hum fades again, and she cannot say when.
That small experiment holds the whole result. The fan’s model in her auditory cortex never changed. It was accurate before she listened and accurate after. What changed was the depth at which its content was handled, the precision given to it, and whether a further model represented it to her as something she was hearing. When all three switched on, she heard the fan. When they switched off, she lost it, and lost the ability to say when she lost it.
So the answer to the question is a conjunction, not a single mechanism. A prediction is experienced when it is held deep, weighted high and represented as one’s own. That answer fits the patients, the sleepers, the hallucinations and the tāla. It does not settle whether the fan was faintly there all along. Every test we have was built by asking people, and the fan is what people do not mention.
The practical decision is narrower and firmer. When an earbud, a chatbot or a brain scan is said to “predict”, ask three questions: at which depth, with what precision, and does it model its own predicting? A yes to the first question alone describes the part of the brain that predicts in sleep.
The fan also leaves a lesson for RTM. Its layers were written as a ladder, with reflexivity above manifestation. Deepa’s experiment suggests the rungs do not come apart so neatly. When she attended, the hum showed up and was taken as hers in the same moment. She could not catch a stage at which the fan was present to her but not yet recognised as present. That is weak evidence, a single invented case, and introspection about timing is poor. Still, it fits the correction made in part five. Vimarśa looks less like a further storey and more like the way the ground floor is lit.
Kiran, meanwhile, has started counting the cycle aloud. Guruji says this will pass with practice. He may be describing a model moving from a level that needs the self-model to one that does not, which is what practice does.
Utpaladeva would say the lover was in the room the whole time. The neuroscience agrees, and adds that being in the room is the easy part.
Sources
Primary research
- Rao, R. P. N. and Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. doi:10.1038/4580
- Bekinschtein, T. A., Dehaene, S., Rohaut, B., Tadel, F., Cohen, L. and Naccache, L. (2009). Neural signature of the conscious processing of auditory regularities. PNAS, 106(5), 1672–1677. doi:10.1073/pnas.0809667106
- Wacongne, C., Labyt, E., van Wassenhove, V., Bekinschtein, T., Naccache, L. and Dehaene, S. (2011). Evidence for a hierarchy of predictions and prediction errors in human cortex. PNAS, 108(51), 20754–20759. doi:10.1073/pnas.1117807108
- Strauss, M., Sitt, J. D., King, J.-R. et al. (2015). Disruption of hierarchical predictive coding during sleep. PNAS, 112(11), E1353-E1362. doi:10.1073/pnas.1501026112
- Melloni, L., Schwiedrzik, C. M., Müller, N., Rodriguez, E. and Singer, W. (2011). Expectations change the signatures and timing of electrophysiological correlates of perceptual awareness. Journal of Neuroscience, 31(4), 1386–1396. jneurosci.org/content/31/4/1386
- Kok, P., Jehee, J. F. M. and de Lange, F. P. (2012). Less is more: expectation sharpens representations in the primary visual cortex. Neuron, 75(2), 265–270. doi:10.1016/j.neuron.2012.04.034
- Pinto, Y., van Gaal, S., de Lange, F. P., Lamme, V. A. F. and Seth, A. K. (2015). Expectations accelerate entry of visual stimuli into awareness. Journal of Vision, 15(8), 13. doi:10.1167/15.8.13
- Stein, T. and Peelen, M. V. (2015). Content-specific expectations enhance stimulus detectability by increasing perceptual sensitivity. Journal of Experimental Psychology: General, 144(6), 1089–1104. doi:10.1037/xge0000109
- Powers, A. R., Mathys, C. and Corlett, P. R. (2017). Pavlovian conditioning-induced hallucinations result from overweighting of perceptual priors. Science, 357(6351), 596–600. doi:10.1126/science.aan3458
- Whyte, C. J. and Smith, R. (2021). The predictive global neuronal workspace: a formal active inference model of visual consciousness. Progress in Neurobiology, 199, 101918. doi:10.1016/j.pneurobio.2020.101918
- Fleming, S. M. (2020). Awareness as inference in a higher-order state space. Neuroscience of Consciousness, 2020(1), niz020. doi:10.1093/nc/niz020
Theory, argument and review
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11, 127–138. doi:10.1038/nrn2787
- Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204. doi:10.1017/S0140525X12000477
- Hohwy, J. (2012). Attention and conscious perception in the hypothesis testing brain. Frontiers in Psychology, 3, 96. doi:10.3389/fpsyg.2012.00096
- Hohwy, J. and Seth, A. (2020). Predictive processing as a systematic basis for identifying the neural correlates of consciousness. Philosophy and the Mind Sciences, 1(II), 3. doi:10.33735/phimisci.2020.II.64
- Seth, A. K. (2014). A predictive processing theory of sensorimotor contingencies. Cognitive Neuroscience, 5(2), 97–118. doi:10.1080/17588928.2013.877880
- Friston, K. (2018). Am I self-conscious? (Or does self-organization entail self-consciousness?) Frontiers in Psychology, 9, 579. doi:10.3389/fpsyg.2018.00579
- Marvan, T. and Havlík, M. (2021). Is predictive processing a theory of perceptual consciousness? New Ideas in Psychology, 61, 100837. doi:10.1016/j.newideapsych.2020.100837
- Corlett, P. R., Horga, G., Fletcher, P. C., Alderson-Day, B., Schmack, K. and Powers, A. R. (2019). Hallucinations and strong priors. Trends in Cognitive Sciences, 23(2), 114–127. doi:10.1016/j.tics.2018.12.001
- Lau, H. and Rosenthal, D. (2011). Empirical support for higher-order theories of conscious awareness. Trends in Cognitive Sciences, 15(8), 365–373. doi:10.1016/j.tics.2011.05.009
- Walsh, K. S., McGovern, D. P., Clark, A. and O’Connell, R. G. (2020). Evaluating the neurophysiological evidence for predictive processing as a model of perception. Annals of the New York Academy of Sciences, 1464(1), 242–268. doi:10.1111/nyas.14321
- Litwin, P. and Miłkowski, M. (2020). Unification by fiat: arrested development of predictive processing. Cognitive Science, 44(7), e12867. doi:10.1111/cogs.12867
- Frankish, K. (2016). Illusionism as a theory of consciousness. Journal of Consciousness Studies, 23(11–12), 11–39. author’s eprint
Primary texts and translations
- Buddhaghosa, Visuddhimagga XIV.130, Pāli with the English of Bhikkhu Ñāṇamoli, The Path of Purification, as transcribed at The Path of Purification blog. To verify against the printed Buddhist Publication Society edition.
- Utpaladeva, Īśvarapratyabhijñākārikā with Vṛtti, verses 1.5.11 and 4.17. GRETIL e-text. Standard edition: R. Torella, The Īśvarapratyabhijñākārikā of Utpaladeva with the Author’s Vṛtti, Motilal Banarsidass, 2002. Translations here are mine.
Secondary and leads
- Parr, T., Corcoran, A. W., Friston, K. J. and Hohwy, J. (2019). Perceptual awareness and active inference. Neuroscience of Consciousness, 2019(1), niz012. doi:10.1093/nc/niz012
- Mahānaya, Īśvarapratyabhijñākārikā with word-by-word English, for the 4.1.17 numbering. mahanaya.org
Cite this work
Mahapatra, R. R. Predictive processing and the distinction between conscious and unconscious inference. Mind Frontier paper 36 (U0877). rajeshrmahapatra.com. DOI: to be assigned.
@article{mahapatra_mf_u0877,
author = {Mahapatra, Rajesh Ranjan},
title = {Predictive processing and the distinction between
conscious and unconscious inference},
journal = {Mind Frontier},
number = {36},
note = {Topic U0877},
url = {https://rajeshrmahapatra.com/papers/mind-frontier/U0877-predictive-processing-conscious-unconscious-inference.html},
doi = {10.xxxx/placeholder}
}