Building a consciousness-detector

Joel Frohlich has an interesting article up at Aeon on the possibility of detecting consciousness.  He begins with striking neurological case studies, such as the one of a woman born without a cerebellum, yet fully conscious, indicating that the cerebellum is not necessary for consciousness.

He works his way to the sobering cases of consciousness detected in patients previously diagnosed as vegetative, accomplished by scanning their brain while asking them to imagine specific scenarios.  He also notes that, alarmingly, consciousness is sometimes found in places no one wants it, such as anesthetized patients.

All of which highlight the clinical need to find a way to detect consciousness, a way independent of behavior.

Frohlich then discusses a couple of theories of consciousness.  Unfortunately one of them is Penrose and Hammeroff’s quantum consciousness microtuble theory.  But at least he dismisses it, citing its inability to explain why the microtubules in the cerebellum don’t make it conscious.  It seems like a bigger problem is explaining why the microtubules in random blood cells don’t make my blood conscious.

Anyway, his preferred theory is integrated information theory (IIT).  Most of you know I’m not a fan of IIT.  I think it identifies important attributes of consciousness (integration, differentiation, causal effects, etc), but not ones that are by themselves sufficient.  It matters what is being integrated and differentiated, and why.  The theory’s narrow focus on these factors, as Scott Aaronson pointed out, leads it to claim consciousness in arbitrary inert systems that very few people see as conscious.

That said, Frohlich does an excellent job explaining IIT, far better than many of its chief proponents.  His explanation reminds me that while I don’t think IIT is the full answer, it could provide insights into detecting whether a particular brain is conscious.

Frohlich discusses how IIT inspired Marcello Massimini to construct his perturbational complexity index, an index used to asses the activity in the brain after it is stimulated using transcranial magnetic stimulation (TMS), essentially sending an electromagnetic pulse through the skull into the brain.  A TMS pulse that leads to the right kind of widespread processing throughout the brain is associated with conscious states.  Stimulation that only leads to local activity, or the wrong kind of activity, isn’t.

IIT advocates often cite the success of this technique as evidence, but from what I’ve read about it, it’s also compatible with the other global theories of consciousness such as global workspace or higher order thought.  It does seem like a challenge for local theories, those that see activity in isolated sensory regions as conscious.

Finally, Frohlich seems less ideological than some IIT advocates, more open to things like AI consciousness, but notes that detecting it in these systems is yet another need for a reliable detector.  I fear detecting it in alternate types of systems represents a whole different challenge, one I doubt IIT will help with.

But maybe I’m missing something?

Postdictive perception and the consciousness finish line

(Warning: neuroscience weeds)

Recently I noted that one of the current debates in cognitive science is between those who see phenomenal and access consciousness as separate things, and those who see them as different aspects of the same thing.  Closely related, perhaps actually identical, is the debate between local and global theories of consciousness.

Diagram of the brain showing sensory and motor cortices
Image credit: via Wikipedia: Blausen.com staff (2014). “Medical gallery of Blausen Medical 2014”. WikiJournal of Medicine 1 (2). DOI:10.15347/wjm/2014.010. ISSN 2002-4436.

Local theories tend to see processing in sensory cortices (visual, auditory, etc) as sufficient for conscious perception of whatever sensory impression they’re processing.  One example is micro-consciousness theory, which holds that processing equals perception.  So formation of a neural image map in sensory cortices equals consciousness of that image.

(A lot of people seem to hold this processing equals perception idea intuitively.  It fits with ideas from people like Jaak Panksepp or Bjorn Merker.)

Another more sophisticated version is Victor Lamme’s local recurrent processing theory, which adds the requirement of recurrent processing in the sensory cortex, processing that involves feed forward signalling as well as feed back signalling in loops.  I discussed local recurrent theory a while back.

Global cognitive theories require that content have large scale effects throughout the brain to become conscious.  Similar to Lamme’s theory, they often see recurrent processing as a prerequisite, but require that it span regions throughout the thalamo-cortical system, either to have wide scale causal effects (global workspace theories) or to reach certain regions (higher order thought theories).

I mentioned that the local vs global debate may be identical to the phenomenal vs access one.  This is primarily because local theories only make sense if phenomenal consciousness is happening in the local sensory cortices independent of access consciousness.  That processing is inherently pre-access, before the content is available for reasoning, action, and report.

The latest shot in this debate is a paper by Matthias Michel and Adrien Doerig, which will be appearing in Mind & Language: A new empirical challenge for local theories of consciousness.

Michel and Doerig look at a type of perception they call “long-lasting postdiction”.  An example of postdiction is when subjects are shown a red disk followed in rapid succession by a green disk 20 ms (milliseconds) later, resulting in perceptual fusion, where the subject perceives a yellow disk.  This is an example of short-lasting postdiction.  In order for the fusion to occur, the red image needs to be processed, and then the green one, and then the two fused.

Short lasting postdiction could represent a problem for micro-consciousness theory, since it results in formation of image maps for each image in rapid succession but not in a way the leads to each being consciously perceived.  (It’s not clear to me this is necessarily true.  I can see maybe the two types of processing simply bleeding into each other.  And see below for one possible response from the micro-consciousness camp.)

Short-lasting postdictions are less of a problem for local recurrent theory, because it takes more time for the recurrent processing to spin up.  (The paper has a quick but fascinating discussion of the time it takes for information to get from the retina to the cortex and then to various regions, along with citations I may have to check out.)

It’s the long-lasting postdictions that are a problem for local recurrence.  The paper discusses images of pairs of vernier lines that are shown to test subjects.  The images are in various positions, changing every 50 ms or so across a period of 170-450 ms, resulting in a perception of the lines moving.

There are variations where one of the vernier lines are skewed slightly, but only on the first image, resulting in the subject perceiving the skew for the entire moving sequence, even though the verniers in the later images are aligned.  Another variation has two images, one early in the sequence and one toward the end, skewed in opposite directions, resulting in the two skewed images being averaged together and the averaged shape being perceived throughout the sequence.

The main takeaway is that the conscious perception of the sequence appears to be formed after the sequence.  Given the relatively lengthy sequence time of up to 450 ms, this is thought to exceed the time it takes for local recurrent processing to happen, and bleeds over into the ignition of global processing, representing a challenge for local recurrent theory.

The authors note that local theorists have two possible outs.  One is to say that, for some reason, the relevant local processing actually doesn’t happen in these sequences.  This would require identifying some other mechanism that unconsciously holds the intermediate images.  The other is to say that the images are phenomenally experienced, but then subsequently replaced in the access stage.  But this would result in phenomenal experience that has no chance of ever being reportable.  (What type of consciousness are we now talking about?)  And it would make local theories extremely difficult to test.

Interestingly, the authors note that the longer time scales may need to be reconciled with the ones in cognitive theories, such as the global neuronal workspace, which identifies 300 ms as a crucial point in conscious perception.  In other words, while this is an issue for local theories, it could be one even for global theories.

All of this reminds me of the phi illusion Daniel Dennett discussed in his book Consciousness Explained.  He describes this illusion in the build up to discussing his own multiple drafts model, a variation of global workspace theory.  Dennett’s variation is that there’s no absolute moment when a perception becomes conscious.  His interpretation of the phi illusion, which seems like another case of postdiction, is that there is no consciousness finish line.  We only recognize that a perception is conscious retroactively, when it achieves “fame in the brain” and influences memory and report.

Anyway, I personally think the main flaw with local theories is that the processing in question is too isolated, too much a fragment of what we think of as an experience, which usually includes at least the sensory perception and a corresponding affective feeling.  The affect part requires that regions far from the sensory cortices become stimulated.  Even if the localists can find an answer to the postdiction issue, I think this broader one will remain.

Unless of course I’m missing something.

The seven attributes of minimal consciousness

Cover of The Evolution of the Sensitive SoulI’m still working my way through Simona Ginsburg and Eva Jablonka’s tome: The Evolution of the Sensitive Soul.  This is the second post of a series on their book.  I’m actually on the last chapter, but that last chapter is close to a hundred pages long, and the book’s prose is dense.  Light reading it isn’t.

Still, it includes a vast overview of the study of consciousness and the mind, not just in contemporary times, but going back to the 19th century and beyond.  For anyone looking for a broad historical overview of the scientific study of the mind, and is willing to put in some work to parse the prose, it’s worth checking out.

As I noted in the first post, G&J aren’t focusing on human level consciousness, that is, higher order metacognitive self awareness and symbolic thought, the “rational soul.”  Similar to the work by Todd Feinberg and Jon Mallatt, their focus is on minimal consciousness, often called “primary consciousness”.  They equate this minimal consciousness with sentience, the ability to  have subjective experiencing (they prefer “experiencing” to just “experience”), which they relate to Aristotle’s “sensitive soul.”

Even having defined this scope however, there remains lots of room for different interpretations.  In an attempt to more precisely define the target of their investigation, they marshal information from contemporary neurobiology and cognitive scientists, along with their theories, to describe seven attributes of minimal consciousness.

  1. Global activity and accessibility.  It’s widely agreed that consciousness is not localized in narrow brain regions.  Although the core ignition and distribution mechanisms might be localized to particular networks, it involves content widely available from disparate brain regions broadcast or made available to the other specialty processes that otherwise work in isolation.
  2. Binding and unification.  The unified nature of conscious perception, such as experiencing the sight of a dog rather than all the constituent sensory components.  Many theories see this being associated with the synchronized firing of neurons in various brain regions, built with recurrent connections between those regions.
  3. Selection, plasticity, learning, and attention.  We are generally conscious of only one thing at a time, or one group of related things.  This involves competition and selection of the winner with the losers inhibited.  It also involves plasticity, which enables learning.
  4. Intentionality (aboutness).  Conscious states are about something, which may be something in the world or the body.  The notion of mental representation is tightly related to this attribute.
  5. Temporal “thickness”.  Neural processing that is quick and fleeting is not conscious.  To be conscious of something requires that the activity be sustained through recurrent feedback loops, both locally and globally.
  6. Values, emotions, goals.  Experience is felt, that is, it has a valence, a sense of good or bad, pleasure or pain, satisfaction or frustration.  These are the attributes that provide motivations, impetus, to a conscious system, that propel it toward certain “attractor” states and away from others.
  7. Embodiment, agency, and a notion of “self”.  The brain is constantly receiving feedback from the body, providing a constant “buzz”, the feeling of existence.  This gives the system a feeling of bodily self.  (Not to be confused with the notion of metacognitive self in human level consciousness.)

G&J refer to this as “the emergentist consensus.”  It seems to pull ideas from global workspace theory, various recurrent loop theories, Damasio’s theories of self and embodiment, and a host of other sources.

It’s important to note that these attributes aren’t free standing independent things.  They interact with and depend on each other.  For example, for a sensory image to be consciously perceived (4), it must achieve (1) global availability by winning (3) selective attention by (2) binding, which results in (5) temporal thickness and strengthens the plasticity aspect of (3).  This process may trigger a reaction which goes through a similar process to achieve (6) value.  All with (7) as a constant underlying hum, subtly (or not so subtly) stacking the deck of what wins (3).

So that’s G&J’s target.  Their goal is to identify functionality, capabilities which demonstrate these attributes in particular species.  Their focus is on learning capabilities, which I’ll go into in the next post.

What do you think about these attributes?  Do they strike you as necessary and sufficient for minimal consciousness, the “sensitive soul”?  Or are they too much, bringing in inessential mechanisms?

The response schema

Several months ago Michael Graziano, and colleagues, attempted a synthesis of three families of scientific theories of consciousness: global workspace theory (GWT), higher order theory (HOT), and his own attention schema theory (AST).

A quick (crudely simplistic) reminder: GWT posits that content becomes conscious when it is globally broadcast throughout the brain, HOT when a higher order representation is formed of a first order representation, and AST when the content becomes the focus of attention and it is included in a model of the brain’s attentional state (the attention schema) for purposes of guiding it.

Graziano equates the global workspace with the culmination of attentional processing, and puts forth the attention schema as an example of a higher order representation, essentially merging GWT and HOT with AST as the binding, and contemplating that the synthesis of these theories approaches a standard model of consciousness.  (A play of words designed to resonate with the standard model of particle physics.)

Graziano’s synthesis has generated a lot of commentary.  In fact, there appears to be an issue of Cognitive Neuropsychology featuring the responses.  (Unfortunately it’s paywalled, although it appears that the first page of every response is public.)  I already highlighted the most prominent response in my post on issues with higher order theories, the one by David Rosenthal, the originator of HOT, who argues that Graziano gets HOT wrong, which appears to be the prevailing sentiment among HOT advocates.

But this post is about Keith Frankish’s response.  Frankish, who is the leading voice of illusionism today, makes the point that, from his perspective, theories of consciousness often have one of two failings.  They either aim too low, explaining just the information processing (a dig perhaps at pure GWT) or too high in attempting to explain phenomenal consciousness as if it actually exists, and he tags HOTs as being in this latter category.

His preferred target is to explain our intuitions about phenomenal consciousness, why we think we have it.  (I actually think explaining why we think we have phenomenal consciousness is explaining phenomenal consciousness, but that’s just my terminological nit with illusionism.)  Frankish thinks that AST gets this just right.

But he sees it as incomplete.  What he sees missing is very similar to the issue I noted in my own post on Graziano’s synthesis: the affective or feeling component.  My own wording at the time was that there should be higher order representations of reflexive reactions.  But I’m going to quote Frankish’s description, because I think it gets at things I’ve struggled to articulate.  (Note: “iconsciousness” is Graziano’s term for access consciousness, as opposed to “mconsciousness” for phenomenal consciousness.):

Suppose that as well as an attention schema, the brain also constructs a response schema—a simplified model of the responses primed by iconsciousness.  When perceptual information enters the global workspace, it becomes available to a range of consumer systems—for memory, decision making, speech, emotional regulation, motor control, and so on. These generate responses of various kinds and strengths, which may themselves enter the global workspace and compete for control of motor systems. Across the suite of consumer systems, a complex multi-dimensional pattern of reactive dispositions will be generated. Now suppose that the brain constructs a simplified, schematic model of this complex pattern. This model, the response schema, might represent the reactive pattern as  a multi-dimensional solid whose axes correspond to various dimensions of response (approach vs retreat, fight vs yield, arousal vs depression, and so on). Attaching  information from the model to the associated perceptual state will have the effect of representing each perceptual episode as having a distinctive but unstructured property which corresponds to the global impact of the stimulus on the subject. If this model also guides our introspective beliefs and reports, then we shall tend to judge and say that our experiences possess an indefinable but potent subjective quality. In the case of pain, for example, attended signals from nociceptors prime a complex set of strong aversive reactions, which the response schema models as a distinctive, negatively valenced global state, which is in turn reported as an ineffable awfulness.

Now, Frankish is an illusionist.  For him, this response schema provides the illusion of phenomenal experience.  My attitude is that it provides part of the content of that experience, which is then incorporated into the experience by the reaction of all the disparate specialty systems, but again that’s terminological.  The idea is that the response schema adds the feeling to the sensory information in the global workspace and becomes part of the overall experience.  It’s why “it feels like something” to process particular sensory or imaginative content.

This seems similar to Joseph LeDoux’s fear schema.  LeDoux’s conception is embedded in an overall HOT framework, whereas Frankish’s is more at home in GWT, but they seem in the same conceptual family, a representation, a schema of lower level reactive processes, used by higher order processes to decide which reflexive reactions to allow and which to inhibit.  It’s the intersection between that lower level and higher level processing that we usually refer to as feelings.

Of course, there is more involved in feelings than just these factors.  For instance, those lower level reflexive reactions also produce physiological changes via unconscious motor signals and hormone releases which alter heart rate, breathing, muscle tension, etc, all of which reverberate back to the brain as interoceptive information, which in turn is incorporated into the response schema, the overall affect, the conscious feeling of the response.  There are also a host of other inputs, such as memory associations.

And it isn’t always the lower level responses causing the higher level response schema to fire.  Sometimes the response schema fires from those other inputs, such as memory associations, which in turn trigger the lower level reactions.  In other words, the activation can go both ways.

So, if this is correct, then the response schema is the higher order description of lower level reflexive reactions.  It is an affect, a conscious feeling (or at least a major component of it).  Admittedly, the idea that a feeling is a data model is extremely counter-intuitive.  But as David Chalmers once noted, any actual explanation of consciousness, other than a magical one, is going to be counter-intuitive.

Similar to the attention schema, the existence of something like the response schema (or more likely: response schemata) seems inevitable, the attention schema for top down control of attention, and the response schema for deciding which reflexes to override, that is, action planning.  The only question is whether these are over simplifications of much more complex realities, and what else might be necessary to complete the picture.

Unless of course I’m missing something.

Islands of awareness

(Warning: neuroscience weeds)

An interesting paper by Tim Bayne, Anil Seth, and Marcello Massimi, which came up in my Twitter stream today, asks whether there can be islands of awareness.

Ordinary consciousness involves ongoing interaction with the environment, receiving sensory information, and producing motor output.  It has a functional role, enabling an organism to deal with novel situations, including dangers and opportunities.

However, it is possible, due to injury or pathology, for some or all of this interaction to be lost, for the consciousness to become isolated.  The authors start out discussing various medical situations where people can become almost completely disconnected on the output end, only able to communicate with the world by the twitching of an eyelid, or even worse, be completely locked in, unable to communicate at all.

It’s also possible to become disconnected on the input side.  Some of these cases are reversible, such as what happens in dreaming.  In that state, we’re generally disconnected from the outside world, although with a strong enough stimulus, that can usually quickly be reversed.  There is also the anesthesia drug ketamine, which apparently provides the same disconnection, but doesn’t always extinguish consciousness, sometimes leading to vivid and terrifying experiences.

But what the authors are most concerned about here are actual islands of awareness, where the system in question is completely isolated from the environment.  They explore three scenarios: ex cranio brains in a nutrient vat, hemispherotomy, and cerebral organoids.

An example of ex cranio brains were the disembodied pig brains which were kept alive in a nutrient delivery system.  The nutrient mix included a neural inhibitor to ensure that the brains wouldn’t regain consciousness, but suppose that inhibitor hadn’t been present?  Could such brains actually be conscious?  A lot of people would say no, that consciousness requires interaction with a body.  But if such a brain showed wide scale organized activity, it might “put pressure” on embodied cognition theories.

A hemispherotomy is sometimes performed on a patient with severe epileptic seizures.  It involves severing the connections between the damaged hemisphere and the other side, as well as its connections with the brainstem, thalamus, and other subcortical structures.  However, a hemispherotomy, unlike a hemispherectomy, leaves the tissue in place, with all of its vascular connections.

Could such a disconnected hemisphere be conscious?  The authors note that, under normal circumstances, without the activating signals coming up from the RAS (reticular activating system) in the brainstem, the activity in the disconnected tissue has very low firing rates, equivalent to a deep dreamless sleep.  But, they ask, what would happen if electrodes were inserted and used to stimulate the hemisphere?  Might it then regain some consciousness?

The authors discuss the role of subcortical regions in consciousness.  It’s well established that they provide crucial support, but what is the nature of that support?  Are they causal, constitutive, or both?  Causal means they just cause awareness in cortical tissue but don’t participate in generating or consuming the content.  Constitutive means they do.

Personally, I think with a disconnected thalamus, the question is somewhat moot.  Such a hemisphere’s ability to communicate with its disparate regions would be heavily compromised.  I tend to doubt any awareness is possible under those conditions.  Only if the subcortical connections were kept intact, with only the RAS disconnected, might it be possible to re-stimulate some form of consciousness.

The third scenario is cerebral organoids.  I did a post on these a few months ago.  The chances that any form of awareness exists in these largely random collection of neural cells is so close to zero that I find worrying about it counter-productive.  There’s simply nothing to indicate these small clumps of neural tissue are organized to have any sensory or affective functionality, and without that, it’s hard to call whatever is happening in them consciousness.  We might as well worry about whether brain tumors or other excised tissue are conscious.

The authors worry that as organoids continue to be developed, there may eventually be issues.  I guess that’s possible, but it only seems to be a significant possibility when enough of the various components of the brain start to be included, which still seems very remote.

The authors mention other possibilities, such as an in utero fetus with some pathology causing it to be completely disconnected.  Given how immature the brain is until well into the third trimester, this only becomes a possibility in the last few months of pregnancy, but it is a possibility.  Would such a system, with no history of sensory input, be conscious in anything like our understanding of the word?

But the most likely scenario is the ex cranio one, such as the pig brains.  It seems inevitable that someone will eventually try that experiment without the inhibitors.  What will it mean if the brains in that scenario do show wide scale organized activity?  The paper discusses the difficulty of detection in these scenarios, and of avoiding false positive and false negatives.

It’s worth noting that the major cognitive theories: global workspace (GWT), integrated information (IIT), and  higher order thought (HOT), are compatible with the ex cranio scenario.  But the more fragmented the tissue is, the less compatible these theories are, although IIT may posit even organoids as having some level of consciousness.  The theory that might be most compatible with small fragmentary islands is recurrent processing theory (RPT).

The paper finishes up by noting that neuroscientific progress increases the chances of producing these islands, and that they may already exist.  They call for careful consideration of the ethical issues involved.

What do you think?  Are conscious hemispheres or organoids more likely than I think?  Or are there solid reasons to conclude that disembodied cognition is impossible?

Stimulating the central lateral thalamus produces consciousness

(Warning: neuroscience weeds)

The thalamus.
Image credit: Wikipedia

A couple of people have asked me about this study, described in numerous popular science articles (such as this one).  A monkey had electrodes installed in its brain that allowed scientists to stimulate parts of its thalamus, the region at the center of the brain which links the cortex to the brainstem and other systems, as well as serves as a relay station for some inter-cortical communication.

Stimulating the monkey, while it was anesthetized, in the central lateral thalamus region caused it to wake up, look around, and reach for things.  Ceasing the stimulation caused the monkey to immediately lose consciousness.  Notably, this region is heavily interconnected with frontal and parietal regions.

Diagram showing the various regions of the thalamus
Image credit: Madhero88 via Wikipedia

Interestingly, stimulating the medial dorsal thalamus, which is heavily connected to the prefrontal cortex, “proved less effective”, and stimulating the central medial thalamus, which projects to the striatum, was also less effective.

In other words, consciousness seemed to be associated with the central lateral thalamus region and its projections to layers in the frontoparietal network.

Diagram showing the regions of the brain
Lobes of the brain
Image credit: BruceBlaus via Wikipedia

One interesting point about this study, is it seems to contradict another study from a year or two ago which ruled out the thalamus as having a role in wakefulness (favoring the basal ganglia instead, if I recall correctly), a reminder that it’s not a good idea to hang too tightly on the results of individual studies.  Another point is the demonstration that the frontoparietal network overall, not just the prefrontal cortex, seemed to be most important for stimulating consciousness.

What does it all mean?  Well, it seems like a dramatic experiment.  And it seems to re-establish the role of the thalamus in wakefulness.  The part about stimulating the regions projecting to the prefrontal cortex not being effective makes me wonder about implications for higher order theories that focus on that region.

All that said, I think we have to bear in mind the distinction between the state of consciousness, that is wakefulness or vigilance, and awareness.  A lot of the information in this experiment seems to be about the state more than awareness.  In that sense, some of the anatomical details are new, but the overall macroscopic picture doesn’t seem to be much affected.

But this is a technical paper and there are probably implications I’m missing.  In particular, the implications for anesthesiology  and other clinical situations may be very significant.

Prefrontal activity associated with the contents of consciousness

The other day I bemoaned the fact that the Templeton competition between global workspace theory (GWT) and integrated information theory (IIT) would take so long, particularly the point about having to wait to see the role of the front and back of the brain in consciousness clarified.  Well, it looks like many aren’t waiting, and studies seem to be piling up showing that the frontal regions have a role.

In a preprint of a new study, the authors discuss how they exposed monkeys to a binocular rivalry type situation.  They monitored the monkeys using a no-report protocol, to minimize the possibility that the monitored activity was more about the need to report than perception.  In this case, the no-report was achieved by monitoring a reflexive eye movement that had been previously shown to correlate with conscious perception.  So the monkeys didn’t have to “report” by pressing a button or any other kind of volitional motor action.

The authors were able “decode the contents of consciousness from prefrontal ensemble activity”.  Importantly, they were able to find this activity when other studies hadn’t, because while those other studies had depended on fMRI scans using blood oxygen levels, this study used equipment physically implanted in the monkey’s brain.

These results add support for cognitive theories of consciousness, such as GWT and higher order theories (HOT), and seem to contradict the predictions made by IIT.

Of course, it doesn’t close off every loophole.  There was speculation on Twitter that Ned Block will likely point out that some variation of his no-post-perceptual-cognition protocol is necessary.  In other words, it can’t be ruled out that the activity wasn’t the monkeys having cognition about their perception after the perception itself.  (Which of course assumes that cognition about the perception and conscious perception are distinct things, something cognitive theories deny.)

And as I’ve noted before, I tend to doubt that the prefrontal cortex’s role will be the whole story, which seems necessary for strict HOT.  It seems possible that someone could have sensory consciousness without it, but probably not affect consciousness, and not introspective consciousness.

So, not the last word, but important results.  After the study last week calling into question the role of the P3b wave, it seems to get global neuronal workspace off the ropes.

Recurrent processing theory and the function of consciousness

Victor Lamme’s recurrent processing theory (RPT) remains on the short list of theories considered plausible by the consciousness science community.  It’s something of a dark horse candidate, without the support of global workspace theory (GWT) or integrated information theory (IIT), but it gets more support among consciousness researchers than among general enthusiasts.  The Michael Cohen study reminded me that I hadn’t really made an effort to understand RPT.  I decided to rectify that this week.

Lamme put out an opinion paper in 2006 that laid out the basics, and a more detailed paper in 2010 (warning: paywall).  But the basic idea is quickly summarized in the SEP article on the neuroscience of consciousness.

RPT posits that processing in sensory regions of the brain are sufficient for conscious experience.  This is so, according to Lamme, even when that processing isn’t accessible for introspection or report.

Lamme points out that requiring report as evidence can be problematic.  He cites the example of split brain patients.  These are patients who’ve had their cerebral hemispheres separated to control severe epileptic seizures.  After the procedure, they’re usually able to function normally.  However careful experiments can show that the hemispheres no longer communicate with each other, and that the left hemisphere isn’t aware of sensory input that goes to the right hemisphere, and vice-verse.

Usually only the left hemisphere has language and can verbally report its experience.  But Lamme asks, do we regard the right hemisphere as conscious?  Most people do.  (Although some scientists, such as Joseph LeDoux, do question whether the right hemisphere is actually conscious due to its lack of language.)

If we do regard the right hemisphere as having conscious experience, then Lamme argues, we should be open to the possibility that other parts of the brain may be as well.  In particular, we should be open to it existing in any region where recurrent processing happens.

Communication in neural networks can be feedforward, or it can include feedback.  Feedforward involves signals coming into the input layer and progressing up through the processing hierarchy one way, going toward the higher order regions.  Feedback processing is in the other direction, higher regions responding with signals back down to the lower regions.  This can lead to a resonance where feedforward signals cause feedback signals which cause new feedforward signals, etc, a loop, or recurrent pattern of signalling.

Lamme identifies four stages of sensory processing that can lead to the ability to report.

  1. The initial stimulus comes in and leads to superficial feedforward processing in the sensory region.  There’s no guarantee the signal gets beyond this stage.  Unattended and brief stimuli, for example, wouldn’t.
  2. The feedforward signal make it beyond the sensory region, sweeping throughout the cortex, reaching even the frontal regions.  This processing is not conscious, but it can lead to unconscious priming.
  3. Superficial or local recurrent processing in the sensory regions.  Higher order parts of these regions respond with feedback signalling and a recurrent process is established.
  4. Widespread recurrent processing throughout the cortex in relation to the stimulus.  This leads to binding of related content and an overall focusing of cortical processes on the stimulus.  This is equivalent to entering the workspace in GWT.

Lamme accepts that stage 4 is a state of consciousness.  But what, he asks, makes it conscious?  He considers that it can either be the location of the processing or the type of processing.  But for the location, he points out that the initial feed forward sweep in stage 2 that reaches widely throughout the brain doesn’t produce conscious experience.

Therefore, it must be the type of processing, the recurrent processing that exists in stages 3 and 4.  But then, why relegate consciousness only to stage 4?  Stage 3 has the same type of processing as stage 4, just in a smaller scope.  If recurrent processing is the necessary and sufficient condition for conscious experience, then that condition can exist in the sensory regions alone.

But what about recurrent processing, in and of itself, makes it conscious?  Lamme’s answer is that synaptic plasticity is greatly enhanced in recurrent processing.  In other words, we’re much more likely to remember something, to be changed by the sensory input, if it reaches a recurrent processing stage.

Lamme also argues from an IIT perspective, pointing out that IIT’s Φ (phi), the calculated quotient of consciousness, would be higher in a recurrent region than in one only doing feedforward processing.  (IIT does see feedback as crucial, but I think this paper was written before later versions of IIT used the Φmax postulate to rule out talk of pieces of the system being conscious.)

Lamme points out that if recurrent processing leads to conscious experience, then that puts consciousness on strong ontological ground, and makes it easy to detect.  Just look for recurrent processing.  Indeed, a big part of Lamme’s argument is that we should stop letting introspection and report define our notions of consciousness and should adopt a neuroscience centered view, one that lets the neuroscience speak for itself rather than cramming it into preconceived psychological notions.

This is an interesting theory, and as usual, when explored in detail, it turns out to be more plausible than it initially sounded.  But, it seems to me, it hinges on how lenient we’re prepared to be in defining consciousness.  Lamme argues for a version of experience that we can’t introspect or know about, except through careful experiment.  For a lot of people, this is simply discussion about the unconscious, or at most, the preconscious.

Lamme’s point is that we can remember this local recurrent processing, albeit briefly, therefore it was conscious.  But this defines consciousness as simply the ability to remember something.  Is that sufficient?  This is a philosophical question rather than an empirical one.

In considering it, I think we should also bear in mind what’s absent.  There’s no affective reaction.  In other words, it doesn’t feel like anything to have this type of processing.  That requires bringing in other regions of the brain which aren’t likely to be elicited until stage 4: the global workspace.  (GWT does allow that it could be elicited through peripheral unconscious propagation, but it’s less likely and takes longer.)

It’s also arguable that considering the sensory regions alone outside of their role in the overall framework is artificial.  Often the function of consciousness is described as enabling learning or planning.  Ryota Kanai, in a blog post discussing his information generation theory of consciousness (which I highlighted a few weeks ago), argues that the function of consciousness is essentially imagination.

These functional descriptions, which often fit our intuitive grasp of what consciousness is about, require participation from the full cortex, in other words, Lamme’s stage 4.  In this sense, it’s not the locations that matter, but what functionality those locations provide, something I think Lamme overlooks in his analysis.

Finally, similar to IIT’s Φ issue, I think tying consciousness only to recurrent processing risks labeling a lot of systems conscious that no one regards as conscious.  For instance, it might require us to see an artificial recurrent neural network as having conscious experience.

But this theory highlights the point I made in the post on the Michael Cohen study, that there is no one finish line for consciousness.  We might be able to talk about a finish line for short term iconic memory (which is largely what RPT is about), another for working memory, one for affective reactions and availability for longer term memory, and perhaps yet another for availability for report.  Stage 4 may quickly enable all of these, but it seems possible for a signal to propagate along the edges and get to some of them.  Whether it becomes conscious seems like something we can only determine retrospectively.

Unless of course I’m missing something?  What do you think of RPT?  Or of Lamme’s points about the problems of relying on introspection and self report?  Should we just let the neuroscience speak for itself?

Stephen Macknik’s work on prosthetic vision

This is pretty wild.  In her latest Brain Science podcast, Ginger Campbell interviews Stephen Macknik on his work to develop a visual replacement implant for blind people.  For a quick overview, check out this short video.

One question Campbell asks, that I was wondering myself: how does the light reach the neurons in the LGN nucleus of the thalamus deep in the center of the brain?  Macknik points out that the gene therapy causes the protein receptor genes to develop in the whole neuron, including the axon terminals that reach into the visual cortex, so the implant doesn’t need to project to the thalamus, just to the end of their axons in the visual cortex.

For more details, check out the podcast itself (it’s about 69 minutes).  It gets a bit technical, but it’s a fascinating interview, and a powerful demonstration about how neuroscientific knowledge can be used.

Is there a conscious perception finish line?

Global workspace theory (GWT) is the proposition that consciousness is composed of contents broadcast throughout the brain.  Various specialty processes compete for the limited capacity of the broadcasting mechanisms, to have their content broadcast to the all the other specialty processes.

Global neuronal workspace (GNW) is a variant of that theory, popularly promoted by Stanislas Dehaene, which I’ve covered before.  GNW is more specific than generic GWT on the physical mechanisms involved.  It relies on empirical work done over the years demonstrating that conscious reportability involves wide scale activation of the cortex.

One of the observed stages is a massive surge about 300 milliseconds after a stimulus, called the P3b wave.  Previous work seemed to establish that the P3b wave is a neural correlate of consciousness.  Dehaene theorized that it represents the stage where one of the signals achieves a threshold and wins domination, with all the other signals being inhibited.  Indeed, the distinguishing mark of the P3b is that it is massively negative in amplitude, indicating that most of it comes from inhibitory action.

The P3b has been replicated extensively and been seen as a pretty established phenomenon associated with attention and consciousness.  But this is science, and any result is always provisional.  Michael Cohen and colleagues have put out a preprint of a study that may demonstrate that the P3b wave is not associated with conscious perception, but with post perceptual processing.

The study tests the perception of subjects, showing various images while measuring their brain waves via EEG.  Using a no-report protocol, in half of the tests, the subjects were asked to report on whether they saw something, but in the other half they were not asked to report.  Crucially, the P3b wave only manifested in the reported cases, never in the non-report ones, even when the non-report image were exactly the same as the ones that did generate affirmative reports.

Image showing P3 wave presence for report and absence for non-report tests
Image from the study: https://www.biorxiv.org/content/10.1101/2020.01.15.908400v1.full

To control for the possibility that the subjects weren’t actually conscious of the image in the non-report cases, the subjects were given a memory test after a batch of non-report events, checking to see what they remember perceiving.  Their memories of the perception correlated with the results in the report versions.

So, the P3b wave, a major piller of GNW, may be knocked down.  The study authors are careful to make clear that this does not invalidate GWT or other cognitive theories of consciousness.  They didn’t test for all the other ways the information may have propagated throughout the cortex.  Strictly speaking, it doesn’t even invalidate GNW itself, but it does seem to knock out a major piece of evidence for it.

However, this is a more interesting discussion if we ask, what would it mean if all cortical communication beyond the sensory regions were ruled out, that the ability to acquire a memory of a sight only required the local sensory cortices?  It might seem like a validation of views like Victor Lamme’s local recurrent processing theory, which holds that local processing in the sensory cortices is sufficient for conscious perception.

But would it be?  Dehaene, when discussing his theory, is clear that it’s a theory of conscious access.  For him, something isn’t conscious until it becomes accessible by the rest of the brain.  Content in sensory cortices may form, but it isn’t conscious until it’s accessible.  Dehaene refers to this content as preconscious.  It isn’t yet conscious, but it has the potential to become so.

In that view, the content of what the subjects perceived in the non-report tests may have been preconscious, unless and until their memories were probed, at which point it became conscious.

This may be another case where the concept of consciousness is causing people to argue about nothing.  If we describe the situation without reference to it, the facts seem clear.

Sensory representations form in the local sensory cortex.  A temporary memory of that representation may persist in that region, so if probed soon enough afterward, a report about the representation can be extracted from it.  But until there is a need for a report or other usage, it is not available to the rest of the system, and none of the activity, including the P3b, normally associated with that kind of access is evident.

This reminds me of Daniel Dennett’s multiple drafts theory (MDT) of consciousness.  MDT is a variant of GWT, but minus the idea that there is any one event where content becomes conscious.  It’s only when the system is probed in certain ways that one of the streams, one of the drafts, become selected, generally one of the ones that has managed to leave its effects throughout the brain, that has achieved “fame in the brain.”

In other words, Dennett denies that there is any one finish line where content that was previously unconscious becomes conscious.  In his view, the search for that line is meaningless.  In that sense, the P3b wave may be a measure of availability, but calling it a measure of consciousness is probably not accurate.  And it’s not accurate to say that the Lamme’s local recurrent processing is conscious, although it’s also not accurate to relegate it completely to the unconscious.  What we can say is that it’s at a particular point in the stream where it may become relevant for behavior, including report.

Maybe this view is too ecumenical and I’m papering over important differences.  But it seems like giving up the idea of one finish line for consciousness turns a lot of theories that look incompatible into models of different aspects of the same overall system.

None of this is to say that GWT or any of its variants might not be invalidated at some point.  These are scientific theories and are always subject to falsification on new data.  But if or when that happens, we should be clear about exactly what is being invalidated.

Unless of course I’m missing something?