Lamme’s recurrent processing theory: local visual feedback loops without prefrontal broadcast
Local feedback inside visual cortex is the best candidate we have for the moment a seen thing shows up, but the evidence proves it is required, not that it is enough, and the experiment that could prove sufficiency has a definite shape that nobody has yet fully built.
TLDR
- Claim: recurrent feedback among visual areas V1 to V4 is probably necessary for seeing a shape, and it is the leading candidate for when the shape appears. Its sufficiency without frontal broadcast is unproven.
- Mechanism: a fast feedforward sweep sorts features in about 100 milliseconds. Higher areas then send signals back down, binding features into a figure on a ground. Masks, anaesthesia and late magnetic pulses all cut this loop.
- Implication: if local recurrence is enough, some experiences are never reported, and report-based science will always miss them. Theories of machine consciousness that count recurrence as an indicator inherit the same open question.
- Boundary: every test of sufficiency must decide whether experience occurred without asking. If recurrence itself becomes the evidence of experience, the theory can no longer lose, which is a cost, not a win.
- Decision: stop treating ‘recurrence present’ as a finding about experience. Run sufficiency audits that pair blocked broadcast with independent no-report markers, and split the RTM layer for access into local appearance and global availability.
Fifty milliseconds in Room 3
It is a little after eleven at night in Room 3 of a small vision lab in Bengaluru. The air conditioner ticks. The only light is a grey monitor refreshing 120 times a second, and the only sound is the click of a chin rest as Arjun, a 24-year-old engineering student who volunteered for ₹500 and a samosa, settles back into position. Behind him sits Meera, a postdoc with a cold coffee and forty hours of equipment time left on a grant that ends in March. She and Arjun are invented. The experiment they are running is not.
On each trial a textured square appears for one frame, about 8 milliseconds, on a background made of the same texture turned through ninety degrees. The square is defined only by the orientation of its lines. Then, after a delay Meera controls, a mask of scrambled texture fills the screen. Arjun presses one key for “square on the left”, another for “square on the right”, and a third for “I saw nothing”.
When the mask follows the square by 25 milliseconds, Arjun presses “nothing” almost every time. When it follows by 120 milliseconds, he is nearly perfect and slightly bored. The trials Meera cares about sit near 50 milliseconds. There Arjun often pauses. Once, he takes off the headphones and says, “Something was there. I couldn’t tell you where.” Then he guesses left, correctly.
Meera’s EEG cap records the trial. The first wave of activity reaches the back of Arjun’s head on every trial, masked or not. What differs on the hesitant trials is what happens a little later, when signals that went up the visual hierarchy start coming back down. On some trials that return is present and on others it is weak. On only some of the trials where it is present does Arjun press a key that says he saw anything.
Meera has to spend her last forty hours on one of two experiments. One adds a magnetic pulse over the back of the head to break the returning signal at a precise moment. The other removes the keypress entirely and tries to read what Arjun saw from his eyes and his brain without asking him. Her advisor thinks the first is safer. Her own bet is the second. Both experiments aim at the question of this paper. is local feedback inside visual cortex enough for Arjun to see the square, before the frontal and parietal areas that let him report it have joined in?
The theory that says yes is Victor Lamme’s recurrent processing theory. My answer, which the rest of the paper earns, is this. Local recurrence is probably necessary for seeing a figure, and it is the best candidate we have for the moment a figure appears. Whether it is sufficient has not been shown by any experiment so far, because every experiment that finds recurrence without broadcast must then decide, without asking, whether experience happened. The consequence is practical. Meera should run the second experiment, with one specific addition, and the rest of us should stop reading “recurrence was present” as a finding about experience.
Two passes through the same cortex
Start with something you have felt. You glance at a crowded shelf and know, before you could say how, that there is a red book somewhere on it. A moment later the red book is simply there, a thing with edges, sitting in front of the other books. Those two moments feel like one. The brain treats them as two.
The first moment is a fast, one-way sweep. Light hits the retina, signals pass through a relay station in the thalamus called the lateral geniculate nucleus, and arrive at primary visual cortex, the strip at the back of the head that neuroscientists call V1. From there the signals climb a ladder of visual areas, V2, V3, V4 and beyond, each tuned to more complex features. Edges become contours, contours become shapes, shapes become categories. In a monkey, the leading edge of this wave reaches high visual areas within roughly 100 milliseconds. Lamme and Pieter Roelfsema called this the feedforward sweep in a review that remains the cleanest statement of the idea, “The distinct modes of vision offered by feedforward and recurrent processing” in Trends in Neurosciences in 2000.
The sweep is fast and very capable. Simon Thorpe and colleagues showed that people can decide whether a briefly flashed photograph contains an animal, and that brain signals separating animal from non-animal scenes appear about 150 milliseconds after the image (Thorpe, Fize and Marlot, 1996). That is barely enough time for one pass up the ladder. So a feedforward pass can do a lot of sorting.
What it cannot do well, on Lamme’s account, is put things together. A neuron in V1 sees a tiny patch of the world. It knows that the lines in its patch run vertically. It does not know whether its patch belongs to the square or to the background. That knowledge exists only higher up, where neurons see larger regions. For the V1 neuron to “know” it is inside the figure, information has to travel back down. When it does, V1 neurons whose patch sits inside the square fire more than neurons whose patch sits on the background, even though the local lines are identical. This extra firing is called figure-ground modulation, and it arrives late, from around 80 to 100 milliseconds after stimulus onset in monkey V1, well after the first response.
Lamme’s claim is that this second, recurrent phase is where seeing happens. In “Towards a true neural stance on consciousness” in 2006 he laid out four stages. Stage 1 is a feedforward sweep confined to visual areas. Stage 2 is feedforward activity that reaches frontal and motor regions and can steer behaviour, still without awareness. Stage 3 is recurrent processing within visual areas. Stage 4 is widespread recurrence that pulls in frontal and parietal cortex. Stage 4 makes a content reportable, rememberable and usable for planning. Stage 3, Lamme argues, already makes it seen.
So the theory makes a sharp cut that most people never make. It separates having an experience from being able to use it. Ned Block named that distinction long before Lamme’s stages, as the difference between phenomenal consciousness, what it is like to see red, and access consciousness, red being available for thought and report. Lamme’s contribution was to tie each side to a different neural process. Local recurrence is the phenomenal side; global recurrence is the access side. When people talk about recurrent processing theory, this mapping is what they mean.
Analogy 1: the canyon echo
Shout into a canyon and the echo returns shaped by the rock. The outgoing shout is the feedforward sweep. The returning echo, carrying information about the canyon’s walls, is recurrence: the same region now holds a signal that has been shaped by larger structure. Only after the echo does the shout “know” the shape of the place it was shouted into. Where the analogy breaks: an echo is a single bounce, and cortical recurrence is a continuing loop in which lower and higher areas keep adjusting each other for tens to hundreds of milliseconds. An echo also carries no new interpretation. The returning cortical signal does, because it encodes which patch belongs to which object.
Arjun’s three waves
Go back to Room 3. On each trial, three things can happen in Arjun’s cortex, in this order. The first is the feedforward wave, which arrives on every trial, masked or not. The second is the local loop, in which V2, V4 and other areas send signals back to V1 and the square separates from its background. The third is what Stanislas Dehaene and Jean-Pierre Changeux call ignition: a sudden, self-sustaining burst that spreads through frontal and parietal cortex and holds the content steady for hundreds of milliseconds. Ignition is what lets Arjun press a key, talk about the trial, and remember it when Meera asks at the end of the block.
Masking changes which waves happen. A mask that arrives at 25 milliseconds does not stop the first wave. It arrives in V1 while the first wave is still climbing, and its own feedforward signal collides with the returning feedback. The local loop is disrupted. Ignition never starts. Arjun says “nothing”. At 120 milliseconds the loop has had time to form, ignition happens, and Arjun reports the square.
At 50 milliseconds, the interesting middle, the loop sometimes forms and is sometimes cut. On some of the trials where it forms, ignition follows and Arjun reports. On others, the loop forms and then fades. Arjun says “nothing”, or “something”, and guesses. The whole dispute is about those fading trials. Recurrent processing theory says Arjun saw the square, briefly, and could not hold on to it. Global workspace theory says the square was processed but never seen. Same trial, same keypress, two incompatible descriptions.
If you chose “the question has no answer”, you are in good company, and Daniel Dennett would have bought you a drink. I don’t think that answer is forced on us yet. It becomes forced only if no evidence other than report could ever bear on trial B. Part three argues that some evidence can, and names it.
Why local recurrence is the right kind of candidate
Lamme gives a positive reason to put experience at the local loop, beyond the trial data. Feedforward processing produces what he calls a set of features. Recurrent processing produces a scene: features bound into objects, objects placed in depth, figures set against grounds. Our visual experience is a scene. We do not experience a list of edge orientations and colour values. The process whose output has the same structure as experience is the better candidate for producing it. The structure of the loop’s output matches the structure of the experience. That is the strongest positive argument for the theory, and it does not depend on any single experiment.
Lamme added a second argument from learning. Recurrent loops between areas are where plasticity of the kind called Hebbian (cells that fire together strengthen their connections) has the most material to work with, because recurrence keeps both ends of a connection active at once. In his 2010 paper “How neuroscience will change our view on consciousness” he proposed that this link with learning is the function of phenomenal experience: a seen scene is one the brain can learn from. That is a hypothesis, and it is a brave one, because it gives experience a job without making the job report.
The structural argument has a weak point that I want on the table early. Matching structure is evidence of relevance, not of identity. The image on a camera sensor after a demosaicing step also has the structure of a scene. Nobody thinks that makes the camera see. Lamme knows this. His answer is that recurrence in a brain is part of a system that also learns, attends and acts, and that recurrence is the ingredient that makes the difference between the brain seeing and not seeing. That is a claim about what is necessary within brains. It is not yet a claim that recurrence alone would suffice anywhere.
Four results that cut the loop
Four experiments carry most of the weight for recurrent processing theory. Each one leaves the feedforward sweep standing and removes the returning signal. Each one finds that seeing goes with it. Taken together they make a strong case that recurrence is required for a figure to be seen. Read carefully, none of them shows that recurrence is enough.
True story 1: the anaesthetised monkeys
In 1998 Lamme, Karl Zipser and Henk Spekreijse recorded from V1 in monkeys looking at texture figures much like Meera’s square. In awake animals, neurons whose small receptive fields fell inside the figure fired more, late in the response, than neurons on the background. Then the team repeated the recordings with the animals under general anaesthesia (ketamine induction, then nitrous oxide and isoflurane). The neurons still responded to the lines in their patch, with the same orientation tuning as before. The figure-ground boost was gone. In their words, the figure-ground activity “was selectively suppressed” while receptive field tuning “remained unaffected”. The feedforward machinery kept working in an animal that was not conscious. The late, contextual part did not.
This is a beautiful result, and it is easy to overread. Anaesthesia does many things at once. It reduces feedback between areas, and it also reduces frontal activity, thalamic drive and much else. The experiment shows that figure-ground modulation and consciousness disappear together under one drug regime. It does not show which loss caused the loss of consciousness, or that restoring local feedback alone would restore it.
The second result came three years later and is closer to Room 3. Hans Supèr, Spekreijse and Lamme trained monkeys to detect a texture figure and recorded from V1 while they did it. On trials where the monkey detected the figure, the late component of the V1 response, beyond about 100 milliseconds, was present. On trials where the monkey missed it, the late component was suppressed, though the stimulus was physically identical. The early response was the same on hits and misses. The paper’s title states the conclusion: “Two distinct modes of sensory processing observed in monkey primary visual cortex”.
Here the monkey is awake and its state is the same on both kinds of trial. That removes the anaesthesia worry. But a new one arrives. A “hit” is a report. The monkey detected the figure and made an eye movement to it. So the late V1 signal goes with reporting. It might be the cause of seeing, or a consequence of attention having already been drawn to the figure by something upstream, or an echo of the decision coming back down. The authors addressed some of this, arguing the signal sat between sensory and decision stages. The deeper problem is that the experiment, like all detection experiments, sorts trials by what the animal did.
True story 2: the phosphenes that stopped moving
In 2001 Alvaro Pascual-Leone and Vincent Walsh used transcranial magnetic stimulation, brief magnetic pulses through the skull that make neurons fire, on human volunteers. A strong pulse over the motion area V5 makes people see a phosphene: a small flash of light that appears to move. The trick was a second, weaker pulse, too weak to make phosphenes itself, delivered over V1. When that V1 pulse came after the V5 pulse, within a window of roughly 5 to 45 milliseconds, most people saw phosphenes that stayed still rather than moving.
Disrupting V1 after the motion signal had already been generated in V5 removed the experience of motion. The motion area had done its work. Without the trip back to V1, the motion did not appear. The paper was titled “Fast backprojections from the motion to the primary visual area necessary for visual awareness”, and the word “necessary” is exactly right.
I like this experiment more than any other in the field, because it bypasses the eyes. There is no stimulus for the feedforward sweep to carry. The content is generated in V5 directly, and the question is only whether it reaches experience. The answer depends on V1, late. That is as close to a causal test of the backward route as anyone has come in humans.
Its limit is narrow and important. It shows that feedback to V1 is necessary for the motion to be seen. It shows nothing about whether that feedback, without the rest of the brain, would be enough. The volunteers’ frontal cortex was working normally throughout. They reported their phosphenes, which requires everything recurrent processing theory calls stage 4.
The fourth result brings the story back to Meera’s exact display. Johannes Fahrenfort, Steven Scholte and Lamme recorded EEG while people detected a texture-defined square, masked or unmasked. With masking, detection fell to chance. The early response over visual cortex, around 110 milliseconds, was the same whether or not the square was masked. The later signatures associated with recurrent processing were absent under masking. In their summary, “feedforward processing is preserved, even when subject performance is at chance”. Masking, the oldest tool in the vision scientist’s kit, appears to work by cutting the loop.
A fifth result that pulls the other way
One more experiment belongs here, because it uses the same logic and lands somewhere less comfortable for the theory. In 2018 Bram van Vugt, Pieter Roelfsema and colleagues recorded from visual and frontal cortex in monkeys trained to report whether a faint stimulus had appeared. Reported stimuli came with strong, sustained frontal activity. Unreported ones produced weak frontal activity that quickly decayed. Information about weak stimuli “could be lost at successive stages en route from the visual to the frontal cortex”, and microstimulating V1 confirmed that signals could fail to propagate. The model that fitted best was one in which stimuli become reportable when they trigger a nonlinear ignition in higher areas.
Roelfsema was Lamme’s co-author on the 2000 review. Here his group finds that the threshold for report is set by propagation and ignition. That is not a refutation of recurrent processing theory, since the theory never claimed report happens without stage 4. It does show that the threshold everybody can measure sits at ignition, and the threshold recurrent processing theory cares about sits somewhere below it, where only indirect signals reach.
Notice what the four results have in common. Each removes recurrence and finds that report goes away. Anaesthesia removes everything. Hits and misses are sorted by report. The phosphene experiment depends on the volunteers saying “moving” or “still”. Masking is scored by detection. Every one of them tests necessity, and every one of them measures experience by report. That is fine for necessity. It is useless for sufficiency, because the case that matters, recurrence with no report, is precisely the case these methods label “unseen”.
Meera understands this. It is why her advisor’s preferred experiment, the magnetic pulse over V1, would make a good paper and leave her question exactly where it was. A pulse that cuts the loop and abolishes seeing would replicate Pascual-Leone and Walsh with texture squares. It would show again that the loop is needed. It would not tell her whether, on the hesitant trials at 50 milliseconds, Arjun saw the square and lost it.
The sufficiency gap
A theory can be right about what is required and wrong about what is enough. Oxygen is required for a fire. It is not enough for one. Recurrent processing theory needs more than the four results of part two. It needs a case in which local recurrence happens, global broadcast does not, and there is still some reason, other than the theory itself, to think the person saw something.
Ned Block made the most influential attempt to supply that reason. His 2007 paper “Consciousness, accessibility, and the mesh between psychology and neuroscience” argued that phenomenal experience overflows access: we see more than we can report. His central evidence came from a 1960 experiment by George Sperling.
The Sperling display
Sperling flashed grids of letters, typically three rows of four, for about 50 milliseconds. Asked to report the whole grid, people managed around four letters and insisted they had seen more. Then he played a tone just after the grid vanished, high for the top row, middle for the middle, low for the bottom. People could now report about three of the four letters in whichever row was cued. Since they could not know in advance which row would be cued, most of the grid must have been briefly available. The report itself was the bottleneck, as described in “The information available in brief visual presentations”.
Lamme’s group extended this into what they called fragile visual short-term memory. Ilja Sligte, Scholte and Lamme found that late cues could still recover more items than ordinary working memory holds. The catch was that no new display could overwrite the first (Sligte, Scholte and Lamme, 2008). The same lab had earlier shown large-capacity storage of integrated objects before change blindness set in (Landman, Spekreijse and Lamme, 2003). The proposal was that this fragile store is what local recurrence holds, and that what it holds is seen.
What this is for: testing your intuition about how much of a brief display is available. What it is not: evidence about whether the letters were experienced.
The sceptic in that card has real names. Michael Cohen and Daniel Dennett argued in 2011 that “Consciousness cannot be separated from function”. Their “perfect experiment” goes like this. Suppose we could find, in someone’s brain, local recurrence for a red patch with no access at all: no report, no memory, no influence on anything. The theory says they experienced red. The person sincerely says they saw nothing. Nothing in their behaviour, now or later, suggests otherwise. What could ever count as evidence that they saw red? If the answer is “the recurrence itself”, then the claim cannot be tested. It has become a definition.
This is the strongest version of the challenge, and I accept its logic. A sufficiency claim that can only be confirmed by its own proposed mechanism is not yet a scientific claim. Where I part company with Cohen and Dennett is the next step. They treat the perfect experiment as the only kind of evidence possible. It is not. Between “full access” and “no consequences at all” there is a wide band of partial consequences, and that band is where the decisive experiment lives.
The attentional blink result
The best evidence so far for sufficiency came from Fahrenfort and colleagues in 2017. They used the attentional blink: when two targets come in quick succession, people often miss the second one because attention is still busy with the first. Here the second target was a set of fragments that, when integrated, formed a surface. Behaviourally, the blink impaired people’s conscious decisions about whether the surface was present. But a classifier trained on EEG could still decode whether an integrated surface had been shown, with the same signature as when people saw it. Masking, by contrast, destroyed both the decision and the decodable signature. The authors concluded that “access consciousness and perceptual integration can be dissociated”.
This is a double dissociation of exactly the right shape. One manipulation, masking, removes integration and access together. Another, the blink, removes access and leaves integration intact. If integration is the local loop and access is ignition, then the blink trials are cases of recurrence without broadcast. The open question is the one the thought experiments keep returning to: were those blinked surfaces seen?
The paper itself is careful about this. It claims that the neural signature of an unaccessed object can be indistinguishable from that of a perceptually rich representation. It does not claim the unaccessed object was experienced. That restraint is correct, and it marks the gap precisely. The signature of a seen object, present without report, is necessary evidence for sufficiency and not yet enough.
What this is for: seeing how two theories read the same toy trial as the mask delay and frontal access change. What it is not: a model fitted to data or evidence about experience. Assumptions: the loop needs about 45 ms to form; ignition needs the loop plus access.
A test built from Lamme’s own hypothesis
Here is my proposal for crossing the gap, and it borrows its force from Lamme. In 2010 he suggested that the function of phenomenal experience is to enable a certain kind of learning, because recurrent loops create the conditions for synapses at both ends of a connection to strengthen. Take that seriously as a prediction. Suppose local recurrence produces experience, and experience enables learning. Then unreported stimuli that received local recurrence should produce learning that stimuli without recurrence do not. That learning should also resemble the learning from seen stimuli, and differ from ordinary unconscious priming.
That gives three groups of trials, all with the same physical stimulus and all with the same “nothing” report. Group one is masked early, so no recurrence. Group two is blinked, so recurrence without broadcast, confirmed on the EEG. Group three is seen and reported, the positive control. A day later, test all three for a kind of learning that ordinary unconscious processing is known not to support well, such as learning a new conjunction of features rather than a single feature. If group two behaves like group three, the theory has earned something no report could have given it. If group two behaves like group one, the theory’s functional story fails, and so does the best route to its sufficiency claim.
The test is not perfect. A defender of global workspace theory can say the blinked stimuli were preconscious, and that preconscious processing supports more learning than we thought. That is a fair reply, and it would itself be informative, because it commits the workspace theorist to a prediction about learning without access. What matters is that the theories now disagree about a measurable outcome a day later, instead of about a label for the same trial.
Analogy 2: the darkroom
Old photographs were made in two steps. Exposure put a latent image on the film. Development in a chemical bath made that image visible. Masking is like opening the darkroom door during exposure. The attentional blink is like developing the photograph and then never printing it. The question of sufficiency becomes: is a developed, unprinted negative a photograph? Where the analogy breaks: a negative can be inspected later by anyone. A visual experience that was never accessed leaves no negative that anyone, including the person who had it, can look at. That asymmetry is the whole problem, and the analogy hides it.
For Meera, part three changes the plan. Her second experiment, the no-report one, needs a third leg. It must measure the local loop directly, so that “recurrence present” is a measurement and not an inference. It must confirm that broadcast was absent, by showing frontal decoding failed. And it must include a later consequence, such as learning, that could separate seen from unseen without asking Arjun anything on the trial itself. Forty hours is tight. It is not impossible if she drops the half of her design that only replicates necessity.
What the rivals say about Arjun
Three rival theories read Arjun’s fading trials in three ways, and each reading is serious. It helps to state them as their best defenders would, because each one gets something right that recurrent processing theory has to answer.
Global neuronal workspace theory grew from Bernard Baars’s work and was developed by Dehaene, Changeux and Lionel Naccache. It says a content becomes conscious when it wins access to a network of long-range neurons in prefrontal and parietal cortex. That network broadcasts it to many specialist systems at once. In 2006 Dehaene and colleagues proposed a three-way taxonomy in “Conscious, preconscious, and subliminal processing”. Subliminal processing is too weak to ignite. Conscious processing ignites. In between sits the preconscious: a stimulus strong enough to ignite that does not, because attention is elsewhere. That middle category is Lamme’s stage 3, wearing a different name. The workspace theorist and the recurrent theorist describe the same brain state and disagree only about whether it is experienced.
What the workspace view gets right is the link to everything we can check. Conscious contents, in every case where we can ask, are the ones that can be reported, reasoned with and remembered. A theory that ties experience to that broadcast stays inside the reach of evidence. The cost is that it must say that the vivid but unreportable trials are simply unseen, and it must say this before any test that could tell.
Higher-order theories, in the versions defended by David Rosenthal, Hakwan Lau and Richard Brown, say a state is conscious when the system represents itself as being in that state. A first-order visual representation of the square, however recurrent, is not enough. The square becomes conscious when some higher-order state, plausibly prefrontal, registers “I am seeing a square”. On this view local recurrence can make a representation better, sharper or more stable. It cannot make it experienced. The view gets right something that recurrent processing theory has trouble with: the difference between seeing and knowing that you see. The cost is that the view must explain why removing large parts of prefrontal cortex does not obviously abolish visual experience, a point argued against higher-order theories by Benjamin Kozuch in 2014.
Integrated information theory, from Giulio Tononi, is the strange ally. It holds that consciousness corresponds to a system’s intrinsic cause-effect structure and predicts that the physical substrate of experience is mainly in a posterior “hot zone”. That puts it on Lamme’s side of the front-back argument. In 2017 Melanie Boly and colleagues argued for the posterior view in “Are the neural correlates of consciousness in the front or in the back of the cerebral cortex?”. Brian Odegaard, Robert Knight and Lau argued the opposite in the companion piece “Should a few null findings falsify prefrontal theories of conscious perception?”. The alliance is uneasy. Integrated information theory cares about the structure of causal connections, not whether signals are currently looping, which leads to the unfolding argument discussed in part six.
What the adversarial collaboration did and did not test
In 2025 the Cogitate consortium published the results of a pre-registered adversarial collaboration between proponents of global workspace and integrated information theories, testing their predictions with fMRI, magnetoencephalography and intracranial recordings in a large sample (Cogitate Consortium, 2025). Across 256 participants, information about conscious content turned up in visual, ventrotemporal and inferior frontal cortex, with sustained responses in occipital and lateral temporal cortex that tracked how long a stimulus stayed on screen. The authors summarised the results as aligning with some predictions of each theory “while substantially challenging key tenets of both theories”. For integrated information theory, the missing sustained synchronisation within posterior cortex was the problem. For workspace theory, it was the general lack of ignition at stimulus offset and the limited prefrontal representation of some conscious dimensions.
Recurrent processing theory was not a party to the collaboration. It would be wrong to claim the results as a win for it. Some of them sit comfortably with it, since conscious content was well represented in the back of the brain. The stimuli, though, were clearly visible faces, objects and letters shown for up to about one and a half seconds and fully reportable. Those are not the fading trials. The adversarial collaboration tested the easy case for every theory, which is the correct first step and not the step that decides this paper’s question.
Mr Dsouza connects to a practical rule I hold regardless of which theory wins. Under real uncertainty about a patient, act as if experience may be present. That rule does not depend on recurrent processing theory. The theory makes it vivid by showing how a coherent scientific view could reach the conclusion that someone is seeing while every behavioural test says otherwise.
What this is for: reading one trial through five frameworks. What it is not: a vote; the frameworks are not equally testable.
With JavaScript off, all five readings:
- Recurrent processing: Local loop formed, ignition failed. Arjun saw the square briefly and could not keep it. His “nothing” is an access failure.
- Global workspace: Content was strong but never ignited. Preconscious, not seen. Arjun’s “nothing” is accurate.
- Higher-order: No state represented Arjun as seeing. Whatever V1 to V4 did, it was not experienced.
- Abhidhamma: A slight object. Eye-consciousness arose, the process reached determining, javana never came. A bare moment of seeing without engagement or recognition.
- Pratyabhijñā: Light coloured by the object without reflexive grasp is like a crystal. Without vimarśa there was no experience to lose.
Analogy 3: the newsroom
A newsroom produces many stories that never run. A reporter files a complete, accurate story at 9 pm. At 11 pm the editor picks the front page and the story is cut. Global workspace theory says only the front page counts as news. Recurrent processing theory says the filed story was news the moment it was written, whether or not anyone read it. Higher-order theory says it became news only when the editor’s log recorded “we have this story”. Where the analogy breaks: a filed story exists on a server and can be read later. An unaccessed experience, if there is one, cannot be retrieved by anyone. The analogy also makes the editor a separate person, while in the brain the “editor” is more loops of the same kind.
One more rival deserves a sentence because it often hides inside the others. Predictive processing, which treats perception as the brain’s best guess about the causes of its input, gives feedback a different job. On that view the downward signals carry predictions and the upward ones carry errors. Recurrence is then not a special consciousness ingredient. It is how all perception works. Paper 36 in this series takes that up directly. For now, note that a predictive theorist can accept every result in part two and still deny that recurrence marks the line between seen and unseen.
Seventeen moments and a crystal
Two Indian traditions spent centuries on a question very close to Meera’s. They did not have EEG. They had disciplined introspection, careful argument and a strong interest in what happens in the instant an object meets the mind. Read on their own terms, they disagree with each other about the fading trials, and the disagreement maps closely onto the one between Lamme and his critics.
The Abhidhamma process of cognition
The Theravāda Abhidhamma, the systematic psychology of the Pāli Buddhist tradition, analyses experience into extremely brief moments of mind. Each moment is a citta (चित्त), a single act of consciousness with its accompanying mental factors. The problem it was trying to solve was not neuroscience. It was soteriological and ethical: where, in the flow of experience, are actions with moral weight generated, and where can attention intervene so that greed and aversion do not take hold?
The standard medieval summary, the Abhidhammattha-saṅgaha attributed to Anuruddha, describes a cognitive process (vīthi, वीथि) at the eye door in up to seventeen moments. I follow the edition and translation by Bhikkhu Bodhi, A Comprehensive Manual of Abhidhamma, chapter IV. A visible form strikes the eye. The background stream of mind, the life-continuum (bhavaṅga, भवङ्ग), vibrates and is arrested. A moment of adverting turns toward the object. Then comes eye-consciousness, the bare moment of seeing. Then receiving, investigating and determining. Then, for a strong object, seven moments of javana (जवन), usually translated “impulsion”: the active phase in which the mind fully engages the object and in which ethically weighty intention arises. Finally two moments of registration may follow before the stream sinks back into the life-continuum.
The part that matters for Arjun is what happens to weaker objects. The text sorts objects at the five sense doors into very great, great, slight and very slight. For a great object the process ends after javana, with no registration. For a slight object, in Bodhi’s rendering, the object is “unable to survive till the arising of the javanas”. Eye-consciousness still arises. The process runs to the determining moment, which repeats two or three times, and then the mind subsides into the life-continuum. For a very slight object there is only a vibration of the life-continuum and no process at all.
The commentarial tradition adds a detail that is almost uncanny. The same manual reports Ledi Sayadaw’s view that distinct recognition of an object does not occur in the bare five-door process at all. It occurs in the mind-door processes that follow it. So the tradition separates three things. There is a bare moment of seeing that occurs even for slight objects. There is full engagement, javana, that requires a stronger object. And there is recognition, which happens later still, in a separate process.
Utpaladeva’s crystal
The Pratyabhijñā or “recognition” school of Kashmir Śaivism, founded by Somānanda and systematised by Utpaladeva in the tenth century, approached the same instant from the opposite direction. Its problem was to show, against Buddhist opponents, that consciousness is not a stream of separate moments but a single self-aware reality that recognises itself in everything that appears. Two terms do most of the work. Prakāśa (प्रकाश) is light, the sheer showing-up of anything. Vimarśa (विमर्श) is reflexive awareness, the light grasping itself as light.
Utpaladeva’s most quoted argument about them is short. It is verse 1.5.11 of the Īśvarapratyabhijñākārikā, numbered as in the GRETIL e-text with his own commentary (Raffaele Torella’s critical edition and translation is the scholarly standard). He writes:
My plain translation is this: they know reflective awareness to be the very nature of light. Otherwise light, even when coloured by objects, would be like a crystal or some other insentient thing. In his own short commentary Utpaladeva adds that without this reflexive grasp, light would have mere transparency (svacchatā) and would lack camatkṛti, a word for the savouring, “ah” quality of experience. A crystal takes on the colour of a red flower placed beside it. It does not see red.
This is a direct challenge to recurrent processing theory, and to RTM’s own way of dividing the layers. On Utpaladeva’s view, manifestation without reflexivity is not experience at all. The local loop in Arjun’s visual cortex may “take on the colour” of the square, binding its features into a figure. If nothing grasps that figure as appearing, it is a crystal, however finely structured. That puts Utpaladeva closer to higher-order theorists than to Lamme, though for different reasons and with a different metaphysics.
Break point map: where the traditions stop matching the science
The Abhidhamma parallel holds for the ordering: bare sensory moment, full engagement, later recognition. It breaks on what counts as conscious, because every citta is conscious by definition, and on timing, because the moments are analytic units, not measured durations. The Pratyabhijñā parallel holds for the distinction between showing and reflexive grasp. It breaks on scale: vimarśa is a property of consciousness as such, and in the end of Śiva, not a module in a brain. Treating vimarśa as “prefrontal cortex” would be exactly the mystical inflation in reverse, a flattening of a metaphysical claim into anatomy. The figure below draws both maps.
What this does to RTM’s second layer
The Recognition Theory of Mind, the working hypothesis of this series, names its second layer “manifestation or access” and labels it prakāśa. Recurrent processing theory exposes that label as ambiguous. Lamme’s whole theory is that manifestation and access are different processes that can come apart. If he is right, a single layer called “manifestation or access” is merging two things.
So I propose to split it, as a refinement under test. L2a, local appearance, is a content that is bound, organised and present within a sensory system, the stage 3 of Lamme’s scheme. L2b, global availability, is the same content made usable by the rest of the system for report, memory and control, Lamme’s stage 4 and the workspace theorist’s ignition. RTM’s third layer, reflexivity (vimarśa), sits above both. That split gives Utpaladeva and Lamme a precise place to disagree. Lamme thinks L2a is enough for experience. Utpaladeva thinks nothing below L3 is experience. Workspace theory thinks L2b is where experience begins.
The split also creates predictions RTM must now carry. If L2a and L2b are separable, there should be manipulations that remove one and spare the other in both directions. Masking removes L2a and therefore L2b. The attentional blink, on Fahrenfort’s evidence, removes L2b and spares L2a. A complete case for the split would also need the reverse: L2b without L2a, global availability of content that was never locally organised. Blindsight, the subject of paper 39 in this series, is the natural place to look, since patients there use visual information without the visual cortex that would build the figure. If no case of L2b without L2a can ever be found, the split is lopsided, and L2a might be better treated as a precondition of L2b than as a separate layer.
Recurrence as a machine indicator
Seventeen researchers, led by Patrick Butlin and Robert Long, published a report in 2023 that turned theories of consciousness into a checklist for AI systems. From recurrent processing theory they took two indicators. The first, called RPT-1, is “input modules using algorithmic recurrence”. The second, RPT-2, is “input modules generating organised, integrated perceptual representations”. The report concluded that no current system was conscious and that there were no obvious technical barriers to building systems that meet the indicators.
Notice the word “algorithmic”. The authors chose it on purpose. Algorithmic recurrence means the same processing is applied repeatedly, with its output fed back as input, whether or not the hardware has physical loops. A network unrolled across time steps with shared weights still has it. The report says plainly that this is “a weak condition that many AI systems already meet”. That choice is sensible for a checklist. It also means RPT-1 inherits the sufficiency gap in a more severe form. In a brain, the gap is between local recurrence and experience. In a machine, it is between any repeated computation and experience.
I think the defender’s reply works for brains and fails for machines, for a reason worth stating. In a brain, we have an independent anchor. We know that, in creatures like us, certain recurrent signatures go with report when report is possible. We can then extend that link, with care, to cases where report is blocked. In a machine there is no such anchor. No machine has ever been the kind of system whose reports we already trust as reports of experience. So a recurrent signature in a machine is a structural fact looking for a theory. It is not yet an indicator of anything.
The sufficiency audit
Here is the instrument this paper contributes. It is a scoring sheet for any study, human, animal or machine, that claims to show local recurrence is enough for experience. It does not decide whether experience occurred. It decides whether a study has the shape that could support a sufficiency claim at all.
The audit has five legs. Leg R asks whether local recurrence was measured directly, for example as figure-ground modulation or late feedback-dependent activity, rather than inferred. Leg B asks whether the absence of global broadcast was measured, for example as failed frontal decoding or absent ignition, rather than assumed from the lack of report. Leg M asks whether there is a no-report marker of experience that was validated where report is possible and that is not simply the recurrence signal measured again. Leg C asks whether the study measured a later consequence, such as learning, that could separate seen from unseen without asking on the trial. Leg N asks whether a negative control, a condition with no local recurrence, produced none of the effects.
Each leg scores 0 (absent), 1 (partial or indirect) or 2 (direct). A study supports a necessity claim if R is at least 1 and N is 2. It supports a sufficiency claim only if R and B are both 2, M is at least 1 and independent of R, C is at least 1, and N is 2. Anything less is evidence that recurrence matters, which is valuable and different.
What this is for: scoring whether a study could support a sufficiency claim. What it is not: a verdict on whether experience occurred.
I ran the audit as a short script on six published studies discussed in this paper, scoring each leg from my reading of the papers. The scores are my judgement, not measurements, and someone else could reasonably move a leg by one point. The output, reproduced exactly below, makes the pattern plain. Five of the six are evidence of necessity or of dissociation. None is evidence of sufficiency.
The audit script and its actual output
CASES = {
"Lamme 1998 anaesthesia": dict(R=2,B=0,M=0,C=0,N=2),
"Super 2001 hits vs misses": dict(R=2,B=0,M=0,C=0,N=2),
"Pascual-Leone 2001 TMS": dict(R=1,B=0,M=0,C=0,N=2),
"Fahrenfort 2007 masking": dict(R=2,B=0,M=0,C=0,N=2),
"Fahrenfort 2017 blink": dict(R=2,B=1,M=0,C=0,N=2),
"van Vugt 2018 threshold": dict(R=1,B=2,M=0,C=0,N=1),
"Positive: blink + learning": dict(R=2,B=2,M=1,C=2,N=2),
"Negative: report-only study": dict(R=0,B=0,M=0,C=0,N=0),
}
def verdict(s):
if s["R"]==2 and s["B"]==2 and s["M"]>=1 and s["C"]>=1 and s["N"]==2:
return "sufficiency-shaped"
if s["R"]==2 and s["B"]>=1 and s["N"]==2:
return "dissociation-shaped"
if s["R"]>=1 and s["N"]==2:
return "necessity-shaped"
return "uninformative"
Lamme 1998 anaesthesia necessity-shaped
Super 2001 hits vs misses necessity-shaped
Pascual-Leone 2001 TMS necessity-shaped
Fahrenfort 2007 masking necessity-shaped
Fahrenfort 2017 blink dissociation-shaped
van Vugt 2018 threshold uninformative
Positive: blink + learning sufficiency-shaped
Negative: report-only study uninformativeThe positive case is the hypothetical study proposed in part three, not a published one. The negative case is a generic study that sorts trials by report and measures nothing else. Van Vugt and colleagues score “uninformative” on this audit only because the audit asks about sufficiency of local recurrence, which their study was not designed to test. On the question it was designed for, the threshold for report, it is strong.
The instrument has assumptions, and they should be visible. It assumes that a no-report marker can be validated in report conditions and transferred. That is contested, and paper 38 in this series takes it apart for binocular rivalry. It assumes that learning is a fair consequence to test, which follows Lamme’s own functional hypothesis but might be wrong. If experience has no function, leg C tests nothing. And it treats “independent of R” as a judgement about whether two measures could be the same signal relabelled. That judgement is where most arguments over a real study would happen.
Worked case 1, minimal: one trial, three numbers
Take a single 50-millisecond trial from Room 3. Meera measures figure-ground modulation in occipital EEG at 0.8 on a scale where unmasked trials average 1.0 and early-masked trials average 0.1. Frontal decoding of square position is at chance. Arjun pressed “nothing”. The numbers are illustrative. The audit scores this single trial R = 2, B = 2, M = 0, C = 0, N not applicable within one trial. Verdict: a candidate case of local appearance without global availability, with no evidence either way about experience. That sentence, and not “Arjun saw it”, is what Meera can put in her lab book.
Worked case 2, realistic: a governance review of a vision model
An invented Bengaluru company, Kestrel Labs, builds a vision system for warehouse robots. Its perception module refines a segmentation map over six recurrent passes, each pass feeding the previous map back with the image. The system separates objects from background, fills in occluded edges and assigns depth order. A client’s ethics board asks whether the module “meets consciousness indicators” and what to do about it.
The honest answer has four parts. First, RPT-1 is met: the module uses algorithmic recurrence. Second, RPT-2 is plausibly met: the output is an organised, integrated scene with figure-ground segregation. Third, the sufficiency audit scores the claim “this module has visual experience” as uninformative, because there is no anchor linking any measurement in this system to experience, no broadcast measure, no no-report marker and no consequence test. Fourth, the board should record that two indicators from one theory are present, that this theory’s sufficiency claim is unproven even in humans, and that the presence of these indicators is a reason to keep watching, not a finding. The board should not commission a welfare review on these grounds alone, and should not publish a statement that the system “is not conscious” either. Both would claim more than anyone knows.
That second case is where this paper changes a decision outside the lab. Indicator checklists are useful because they force specificity. They become harmful when a single indicator, met by a common architectural feature, is reported as if it were evidence about experience. My bet is that within a few years, governance documents will routinely list RPT-1 as “met” for most perception systems. If they do, the word “met” should come with the sufficiency audit’s verdict beside it.
The case that recurrence is not experience
Here is the objection in the form its best defenders would give it. I think it is strong, and parts of it are right.
Start with the evidence. Every result that links local recurrence to seeing was obtained with report, or with a stand-in validated against report. When report is removed, the link between recurrence and experience is no longer observed. It is assumed. Recurrent processing theory then reads every case of recurrence without report as a case of unreported experience. Inattentional blindness, the attentional blink and fragile memory are all absorbed this way. A theory that can absorb every case in which its marker appears without its target has stopped making predictions about the target.
Next, the mechanism. Recurrence is everywhere in the brain. It is present in sleep, in anaesthesia at reduced levels, in motor control and in the cerebellum’s loops with the cortex, none of which anyone regards as visual experience. If local recurrence were sufficient, we would need an extra condition saying which recurrence counts. Once that condition is added, it may be the condition, and not the recurrence, that does the work. Lamme’s answer is that the relevant recurrence is the kind that builds a perceptual scene in sensory cortex. That narrows the claim, and it also makes it look more like a description of what seen content is than an explanation of why it is seen.
Then the alternatives. Workspace theory explains every result in part two without positing unreported experience. Under masking, broadcast fails because recurrence fails, and recurrence is necessary for a content strong enough to ignite. Under the blink, recurrence succeeds and broadcast fails, so the content is preconscious. The workspace account is simpler in one clear sense: it posits no experiences that no one could ever have reason to believe in. Higher-order theories add that seeing requires a representation of oneself as seeing, and that nothing in V1 to V4 provides one. Utpaladeva, from a different world entirely, says the same thing in his own terms.
Finally, the machines. The unfolding argument shows that if recurrence is defined by wiring, it is invisible to behaviour, and if it is defined by algorithm, it is too common to matter. The theory has to choose which, and both choices are expensive.
What survives? I think three things. The necessity results survive: removing local recurrence removes seeing, in every paradigm where we can check. The dissociation survives: Fahrenfort’s blink result shows integration and access can be separated. And the structural argument survives: the output of local recurrence has the form of a visual scene, and feedforward output does not. What does not survive, in my judgement, is any claim that sufficiency has been shown. The theory is a strong account of what seeing requires and an unproven account of what seeing is.
What would change my mind
My central claim is that local recurrence is probably necessary, and not yet shown sufficient, for visual experience. These observations would move me.
A clean case of seeing without local recurrence would weaken the necessity half. That would be a person who reliably reports and describes a figure on a ground, with confidence, under conditions where direct recording shows no feedback-dependent activity in early visual cortex. If such cases appeared, recurrence would be an ordinary correlate rather than a requirement, and RTM’s proposed layer L2a would lose its claim to be a precondition of L2b.
A sufficiency-shaped study with a positive result would strengthen the theory a great deal. That means a study that passes all five legs of the audit, with blinked trials producing next-day learning that matches seen trials and differs from masked ones. It would support the RTM split between L2a and L2b and would be evidence against Utpaladeva’s claim, read as an empirical claim, that nothing below reflexivity counts as experience.
The same study with a negative result would move me the other way. If blinked trials with confirmed local recurrence produce learning no different from masked trials, Lamme’s functional story fails. I would then think stage 3 is better described as preparation for experience, and I would fold L2a back into L2 as a precondition rather than a layer.
A demonstration that no-report markers cannot be validated independently of the theories they are used to test would change the field, not just my mind. If every candidate marker turns out to be recurrence measured another way, the sufficiency question becomes unanswerable by any method we have, and Cohen and Dennett win the argument by default.
What I still don’t know
I don’t know whether there is a learning test that unconscious processing provably cannot pass, which leg C needs. I don’t know how much frontal activity is enough to count as broadcast, which leg B needs. I don’t know whether the fragile memory results reflect seen items or partly processed ones. And I don’t know what, if anything, the theory implies for systems whose recurrence is algorithmic rather than physical, beyond the warning in part six.
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 | Visual processing has a fast feedforward phase and a later recurrent phase that can be separated experimentally. | Established | Published evidence | Evidence that late V1 modulation arises without any feedback from higher areas. |
| C2 | Disrupting local recurrence (masking, anaesthesia, late TMS) abolishes reported seeing while sparing feedforward responses. | Established | Published evidence | Reliable reported seeing with feedback-dependent activity directly shown absent. |
| C3 | Local recurrence is necessary for seeing a figure on a ground. | Likely | Published evidence, inference | As C2, or figure perception in patients lacking early-cortex feedback. |
| C4 | Perceptual integration and conscious access can be dissociated. | Likely | Published evidence (single key study) | Failed replication of the attentional-blink decoding result. |
| C5 | No published study has yet shown local recurrence to be sufficient for experience. | Likely | Author’s reading of the literature | A published study passing all five legs of the sufficiency audit. |
| C6 | Report-based methods cannot test sufficiency, since the critical case is scored as unseen. | Established | Argument | A report-based design whose critical trials are not sorted by report. |
| C7 | Unreported but recurrent stimuli will support feature-conjunction learning like seen stimuli. | Speculative | Open hypothesis | The three-group learning study finding blinked trials equal to masked trials. |
| C8 | RTM’s layer L2 should be split into local appearance (L2a) and global availability (L2b). | Speculative | Theoretical refinement | No dissociation of L2b without L2a is ever found, or C7 fails. |
| C9 | Algorithmic recurrence (RPT-1) in an AI system is not, alone, evidence about experience. | Likely | Argument | An independent anchor linking machine recurrence to experience. |
| C10 | Abhidhamma and Pratyabhijñā analyses disagree about whether bare appearance counts as experience, in a way that maps onto the RPT dispute. | Likely | Textual reading | A standard commentary showing the slight-object process lacks eye-consciousness, or that Utpaladeva allows prakāśa without vimarśa to be sentient. |
One dated prediction joins the programme’s Prediction Registry. By 31 December 2030, no peer-reviewed study will meet all five legs of the sufficiency audit for local recurrence in humans. It will be judged against the audit criteria in this paper, applied to studies indexed in PubMed with “recurrent processing” and “consciousness” in the title or abstract. I would be glad to lose this one.
Glossary
- Attentional blink
- A brief period after spotting one target in which people often miss a second target, because attention is still occupied.
- Backward masking
- Showing a second pattern shortly after a first, which can make the first invisible even though it was physically shown.
- Bhavaṅga (भवङ्ग)
- Pāli, “life-continuum”: in the Abhidhamma, the background stream of mind between active cognitive processes.
- Citta (चित्त)
- Pāli and Sanskrit, a single moment or act of consciousness in the Abhidhamma analysis.
- Feedforward sweep
- The first, one-directional wave of activity from the eye up through the visual areas, within roughly 100 milliseconds.
- Figure-ground modulation
- Extra firing in early visual neurons when their small patch lies inside a figure rather than the background, arriving after the first response and dependent on feedback.
- Global availability (RTM L2b)
- This paper’s proposed name for content made usable by the whole system for report, memory and control.
- Ignition
- In workspace theory, a sudden self-sustaining burst of activity across frontal and parietal cortex that makes a content conscious and reportable.
- Javana (जवन)
- Pāli, “impulsion”: the active phase of a cognitive process in which the mind fully engages its object.
- Local appearance (RTM L2a)
- This paper’s proposed name for content bound and organised within a sensory system, before global availability.
- No-report paradigm
- An experiment that infers what someone perceives from signals such as eye movements or pupil size, without asking them.
- Phenomenal and access consciousness
- Block’s distinction between what an experience is like and the experience being available for thought and report.
- Prakāśa (प्रकाश)
- Sanskrit, “light”: in Pratyabhijñā, the sheer showing-up of anything in consciousness.
- Recurrent processing
- Signals passing back from higher to lower areas, and around again, so that a region’s activity reflects the wider context.
- Sufficiency audit
- The five-leg scoring sheet in this paper for whether a study has the shape needed to support a sufficiency claim.
- V1 to V4
- Successive visual areas at the back of the brain. V1 receives input from the thalamus and analyses small patches. V4 handles more complex shape and colour.
- Vimarśa (विमर्श)
- Sanskrit, “reflexive awareness”: in Pratyabhijñā, light grasping itself as light.
- Vīthi (वीथि)
- Pāli, “process” or “street”: the ordered sequence of mind-moments in one act of cognition.
Meera’s last forty hours
Meera does not run the magnetic pulse study. She writes a one-page memo to her advisor that says the pulse study would show, with texture squares, what Pascual-Leone and Walsh showed with phosphenes in 2001. It would be publishable, and it would leave the hesitant trials exactly as mysterious as they are now.
Instead she uses twenty hours for the no-report design with three groups: early-masked, blinked and seen. She measures figure-ground modulation directly over occipital cortex, so leg R is a measurement. She trains a frontal decoder on reported trials and checks that it fails on blinked ones, so leg B is a measurement too. She uses the other twenty hours, spread over the next day, for a learning test on new feature conjunctions. Arjun comes back for a second samosa.
The study might fail. Blinked trials might teach Arjun nothing more than masked ones. If so, Lamme’s functional story is in trouble, and RTM’s proposed split of its second layer goes back in the box. If blinked trials teach like seen ones, the field has something it has never had: a reason, other than recurrence itself and other than report, to think a fading trial was seen.
That is the decision this paper changes, and it applies beyond Room 3. Treat “recurrence present” as a measurement of local appearance, never as a finding about experience. When a study, a patient chart or an AI governance review says recurrence was present, ask which legs of the audit were measured. Most of the time the answer will be one or two. That is still useful. It is just not the answer to the question.
One last thing, which I did not expect when I started. The tradition that came closest to Lamme was not the one that shares RTM’s vocabulary. The Abhidhamma, with its slight objects that reach eye-consciousness and never reach impulsion, already had a place for an appearance that does not get used. Utpaladeva, whose terms give RTM its layer names, would reject that place outright. A theory that borrows his words owes him the honesty of saying that, on this question, he and Lamme cannot both be right.
Sources
Primary research
- Lamme, V. A. F. and Roelfsema, P. R. (2000). The distinct modes of vision offered by feedforward and recurrent processing. Trends in Neurosciences, 23(11), 571–579. doi:10.1016/S0166-2236(00)01657-X
- Lamme, V. A. F. (2006). Towards a true neural stance on consciousness. Trends in Cognitive Sciences, 10(11), 494–501. doi:10.1016/j.tics.2006.09.001
- Lamme, V. A. F. (2010). How neuroscience will change our view on consciousness. Cognitive Neuroscience, 1(3), 204–220. doi:10.1080/17588921003731586
- Lamme, V. A. F., Zipser, K. and Spekreijse, H. (1998). Figure-ground activity in primary visual cortex is suppressed by anesthesia. PNAS, 95(6), 3263–3268. doi:10.1073/pnas.95.6.3263
- Supèr, H., Spekreijse, H. and Lamme, V. A. F. (2001). Two distinct modes of sensory processing observed in monkey primary visual cortex (V1). Nature Neuroscience, 4(3), 304–310. doi:10.1038/89551
- Pascual-Leone, A. and Walsh, V. (2001). Fast backprojections from the motion to the primary visual area necessary for visual awareness. Science, 292(5516), 510–512. doi:10.1126/science.1057099
- Fahrenfort, J. J., Scholte, H. S. and Lamme, V. A. F. (2007). Masking disrupts reentrant processing in human visual cortex. Journal of Cognitive Neuroscience, 19(9), 1488–1497. doi:10.1162/jocn.2007.19.9.1488
- Fahrenfort, J. J., van Leeuwen, J., Olivers, C. N. L. and Hogendoorn, H. (2017). Perceptual integration without conscious access. PNAS, 114(14), 3744–3749. doi:10.1073/pnas.1617268114
- van Vugt, B., Dagnino, B., Vartak, D., Safaai, H., Panzeri, S., Dehaene, S. and Roelfsema, P. R. (2018). The threshold for conscious report: signal loss and response bias in visual and frontal cortex. Science, 360(6388), 537–542. doi:10.1126/science.aar7186
- Thorpe, S., Fize, D. and Marlot, C. (1996). Speed of processing in the human visual system. Nature, 381, 520–522. doi:10.1038/381520a0
- Sperling, G. (1960). The information available in brief visual presentations. Psychological Monographs, 74(11), 1–29. doi:10.1037/h0093759
- Sligte, I. G., Scholte, H. S. and Lamme, V. A. F. (2008). Are there multiple visual short-term memory stores? PLoS ONE, 3(2), e1699. doi:10.1371/journal.pone.0001699
- Landman, R., Spekreijse, H. and Lamme, V. A. F. (2003). Large capacity storage of integrated objects before change blindness. Vision Research, 43(2), 149–164. doi:10.1016/S0042-6989(02)00402-9
- Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. doi:10.1038/s41586-025-08888-1; abstract at PubMed 40307561
Theory, argument and review
- Block, N. (2007). Consciousness, accessibility, and the mesh between psychology and neuroscience. Behavioral and Brain Sciences, 30(5–6), 481–499. doi:10.1017/S0140525X07002786
- Cohen, M. A. and Dennett, D. C. (2011). Consciousness cannot be separated from function. Trends in Cognitive Sciences, 15(8), 358–364. doi:10.1016/j.tics.2011.06.008
- Dehaene, S., Changeux, J.-P., Naccache, L., Sackur, J. and Sergent, C. (2006). Conscious, preconscious, and subliminal processing: a testable taxonomy. Trends in Cognitive Sciences, 10(5), 204–211. doi:10.1016/j.tics.2006.03.007
- Boly, M. et al. (2017). Are the neural correlates of consciousness in the front or in the back of the cerebral cortex? Journal of Neuroscience, 37(40), 9603–9613. doi:10.1523/JNEUROSCI.3218-16.2017
- Odegaard, B., Knight, R. T. and Lau, H. (2017). Should a few null findings falsify prefrontal theories of conscious perception? Journal of Neuroscience, 37(40), 9593–9602. doi:10.1523/JNEUROSCI.3217-16.2017
- Kozuch, B. (2014). Prefrontal lesion evidence against higher-order theories of consciousness. Philosophical Studies, 167, 721–746. doi:10.1007/s11098-013-0109-z
- Bayne, T., Seth, A. K. and Massimini, M. (2020). Are there islands of awareness? Trends in Neurosciences, 43(1), 6–16. doi:10.1016/j.tins.2019.11.003
- Doerig, A., Schurger, A., Hess, K. and Herzog, M. H. (2019). The unfolding argument: why IIT and other causal structure theories cannot explain consciousness. Consciousness and Cognition, 72, 49–59. doi:10.1016/j.concog.2019.04.002
- Butlin, P., Long, R. et al. (2023). Consciousness in artificial intelligence: insights from the science of consciousness. arXiv:2308.08708. arxiv.org/abs/2308.08708
Primary texts and translations
- Anuruddha, Abhidhammattha-saṅgaha, chapter IV (vīthisaṅgaha). Edited and translated in Bhikkhu Bodhi (general editor), A Comprehensive Manual of Abhidhamma, Buddhist Publication Society, 1993.
- Utpaladeva, Īśvarapratyabhijñākārikā with the author’s Vṛtti, verse 1.5.11. GRETIL e-text. Standard critical edition and translation: R. Torella, The Īśvarapratyabhijñākārikā of Utpaladeva with the Author’s Vṛtti, Motilal Banarsidass, corrected edition 2002. Translations in this paper are mine.
Secondary and leads
- Nature Reviews Neuroscience research highlight on Pascual-Leone and Walsh (2001), “Moving pictures”, used for the 5 to 45 millisecond window. nature.com/articles/35077517
- Butlin et al. (2023), HTML version, for the quoted wording on algorithmic recurrence. arxiv.org/html/2308.08708v1
Cite this work
Mahapatra, R. R. Lamme’s recurrent processing theory: local visual feedback loops without prefrontal broadcast. Mind Frontier paper 35 (U0853). rajeshrmahapatra.com. DOI: to be assigned.
@article{mahapatra_mf_u0853,
author = {Mahapatra, Rajesh Ranjan},
title = {Lamme's recurrent processing theory: local visual feedback
loops without prefrontal broadcast},
journal = {Mind Frontier},
number = {35},
note = {Topic U0853},
url = {https://rajeshrmahapatra.com/papers/mind-frontier/U0853-lamme-recurrent-processing-theory.html},
doi = {10.xxxx/placeholder}
}