The issues with higher order theories of consciousness

After the global workspace theory (GWT) post, someone asked me if I’m now down on higher order theories (HOT).  It’s fair to say I’m less enthusiastic about them than I used to be.  They still might describe important components of consciousness, but the stronger assertion that they provide the primary explanation now seems dubious.

A quick reminder.  GWT posits that conscious content is information that has made it into a global workspace, that is, content that has won the competition and, for a short time, is globally shared throughout the brain, either exclusively or with only a few other coherently compatible concepts.  It becomes conscious by all the various processes collectively having access to it and each adapting to it within their narrow scope.

HOT on the other hand, posits that conscious content is information for which a higher order thought of some type has been formed for it.  In most HOTs, the higher order processing is thought to happen primarily in the prefrontal cortex.

As the paper I highlighted a while back covered, there are actually numerous theories out there under the higher order banner.  But it seems like they fall into two broad camps.

In the first are versions which say that a perception is not conscious unless there is a higher order representation of that perception.  Crucially in this camp, the entire conscious perception is in the higher order version.  If, due to some injury or pathology, the higher order representation were to be different than the lower order one, most advocates of these theories say that its the higher order one we’d be conscious of, even if the lower order one was missing entirely.

Even prior to reading up on GWT, I had a couple of issues with this version of HOT.  My first concern is that it seems computationally expensive and redundant.  Why would the nervous system evolve to have formed the same imagery twice?  We know neural processing is metabolically expensive.  It seems unlikely evolution would have settled on such an arrangement, at least unless there was substantial value to it, which hasn’t been demonstrated yet.

It also raises an interesting question.  If we can be conscious of a higher order representation without the lower order one, why then, from an explanatory strategy point of view, do we need the lower order one?  In other words, why do we need the two tier system if one (the higher tier) is sufficient?  Why not just have one sufficient tier, the lower order one?

The HOTs I found more plausible were in the second camp, and are often referred to as dispositional or dual content theories.  In these theories, the higher order thought or representation doesn’t completely replace the lower order one.  It just adds additional elements.  This has the benefit of making the redundancy issue disappear.  In this version, most of the conscious perception comes from the lower order representations, with the higher order ones adding feelings or judgments.  This content becomes conscious by its availability to the higher order processing regions.

But this then raises another question.  What about the higher order region makes it conscious?  By making the region itself, the location, the crucial factor, we find ourselves skirting with Cartesian materialism, physical dualism, the idea that consciousness happens in a relatively small portion of the brain.  (Other versions of this type of thinking locate consciousness in various locations such as the brainstem, thalamus, or hippocampus.)

The issue here is that we still face the same problem we had when considering the whole brain.  What about the processing of that particular region makes it a conscious audience?  Only, since now we’re dealing with a subset of the brain, the challenge is tougher, because it has to be solved with less substrate.  (A lot less with many of the other versions.  At least the prefrontal cortex in humans is relatively vast.)

We can get around this issue by positing that the higher order regions make their results available back into the global workspace, that is, by making the entire brain the audience.  It’s not the higher order region itself which is conscious.  Its contents become conscious by being made accessible to the vast collection of unconscious processes throughout the brain, each of which act on it in its own manner, collectively making that content conscious.

But now we’re back to consciousness involving the workspace and its audience processes.  HOT has dissolved into simply being part of the overall GWT framework.  In other words, we don’t need it, at least not as a theory, in and of itself, that explains consciousness.

None of this is to say higher order processing isn’t a major part of human consciousness.  Michael Graziano’s attention schema theory, for instance, might well still have a role to play in providing top down control of attention, and providing our intuitive sense of how it works.  The other higher order processes provide metacognition, imagination, and what Baars calls “feelings of knowing,” among many other things.

They’re just not the sole domain of consciousness.  If many of them were knocked out, the resulting system would still be able to have experiences, experiences that could lay down new memories.  It’s just that the experience would be simpler, less rich.

Finally, it’s worth noting that David Rosenthal, the original author of HOT, makes this point in response to Michael Graziano’s attempted synthesis of HOT, GWT, and his own attention schema theory (AST):

Graziano and colleagues see this synthesis as supporting their claim that AST, GWT, and HO theory “should not be viewed as rivals, but as partial perspectives on a deeper mechanism.” But the HO theory that figures in this synthesis only nominally resembles contemporary HO theories of consciousness. Those theories rely not on an internal model of information processing, but on our awareness of psychological states that we naturally classify as conscious. HO theories rely on what I have called (2005) the transitivity principle, which holds that a psychological state is conscious only if one is in some suitable way aware of that state.

This implies that consciousness is introspection.  Admittedly, there is precedent going back to John Locke for defining consciousness as introspection.  (Locke’s specific definition was “the perception of what passes in a man’s own mind”.)  Doing so dramatically reduces the number of species that we consider to be conscious, perhaps down to just humans, non-infant humans to be precise.  I toyed with this definition a few years ago, before deciding that it doesn’t fit most people’s intuitions.  (And when it comes to definitions of consciousness, our intuitions are really all we have.)

It ignores the fact that we are often not introspecting while we’re conscious.  And much of what we introspect goes on in animals (in varying degrees depending on species), or human babies for that matter, even if they themselves can’t introspect it.  It also ignores the fact that if a human, through brain injury or pathology, loses the ability to introspect, but still shows an awareness of their world, we’re going to regard them as conscious.

So HOT doesn’t hold the appeal for me it did throughout much of 2019.  Although new empirical results could always change that in the future.

What do you think?  Am I missing benefits of HOT?  Or issues with GWT?

Joseph LeDoux’s theories on consciousness and emotions

The cover of 'The Deep History of Ourselves'In the last post, I mentioned that I was reading Joseph LeDoux’s new book, The Deep History of Ourselves: The Four-Billion-Year Story of How We Got Conscious Brains.  There’s a lot of interesting stuff in this book.  As its title implies, it starts early in evolution, providing a lot of information on early life, although I didn’t find that the latter parts of the book, focused on consciousness and emotion, made much use of the information from the early chapters on evolution.  Still, it was fascinating reading and I learned a lot.

In the Nautilus piece I shared before, LeDoux expressed some skepticism about animal consciousness being like ours.  That seems to be a somewhat milder stance compared to the one in the book.  Here, LeDoux seems, at best, on the skeptical side of agnostic toward non-human animal consciousness.  The only evidence for consciousness he sees as unequivocal is self report, which of course only humans can provide.

In terms of consciousness theories, LeDoux regards Higher Order Theories (HOT) and Global Workspace Theories (GWT) as the most promising, but his money is on HOT, and he provides his own proposed theoretical extensions to it.  HOT posits that consciousness doesn’t lie in the first order representations made in early sensory regions, but in later stage representations that are about these first order ones.  In essence, to be conscious of a representation requires another higher order representation.

In typical HOT, these higher order representations are thought to be in the prefrontal cortex.  LeDoux attributes a lot of functionality to the prefrontal cortex, more than most neuroscientists.  Some of what he attributes I’ve more commonly seen attributed to regions like the parietal cortex.  But he presents information on the connections between various cortical and subcortical regions to the prefrontal cortex to back up his positions.

In the last post, I laid out the hierarchy I usually use to think of cognitive capabilities.  LeDoux has a similar hierarchy, which he discusses in a paper available online, although his is focused on types of behavior.  Going from simpler to more sophisticated:

  1. Species typical innate behavior
    1. Reflexes: Relatively simply survival circuits, centered on the brainstem regions
    2. Fixed Reaction Patterns: More complex survival circuits, often going through subcortical regions such as the amygdala
  2. Instrumental learned behavior
    1. Habits: Actions that persist despite lack of evidence of a good or bad consequence
    2. Action-outcome behaviors: Actions based on the remembered outcomes of past trial-and-error learning
    3. Nonconscious deliberative actions: Actions taken based on prospective predictions made on future outcomes
    4. Conscious deliberative actions: Deliberative actions accompanied and informed by conscious feeling states

On first review, I was unsure about the distinction between action-outcome and deliberative action.  Action-outcome seems like simply a less sophisticated version of deliberative action, particularly since episodic memory and imagined future scenarios are reputed to use the same neural machinery.  It seemed like just different degrees of what I normally label as imaginative planning.

But on further consideration, I can see a case that simply remembering a past pattern of activity and recognizing the same sequence, is not the same thing as simulating new hypothetical scenarios, specific scenarios that the animal has never experienced before.  Put another way, deliberative actions require taking multiple past scenarios and combining them in creative new ways.

Anyway, LeDoux states that there is “no convincing” evidence for instrumental behavior in pre-mammalian vertebrates, or in invertebrates.  In his view, instrumental behavior only exists in mammals and birds.

(This seems to contrast sharply with Feinberg and Mallatt in The Ancient Origins of Consciousness, who cite numerous studies showing instrumental learning in fish, amphibians, and reptiles.  One of the things I’m not wild about LeDoux’s book, is that while he has bibliographic notes, they’re not in-body citations,  making it very difficult to review the sources of his conclusions.)

Deliberative action, on the other hand, LeDoux only sees existing in primates, with humans taking it to a new level.  Apparently in this hierarchy, consciousness only comes into the picture with the most sophisticated version.  I think “consciousness” in this particular context means autonoetic consciousness, that is, introspective self awareness with episodic memory.

(Endel Tulving, the scientist who proposed the concept of autonoesis, doesn’t see episodic memory developing until humans.  However, there is compelling behavioral evidence that it developed much earlier, and is in, at least, all mammals and birds,  although it’s admittedly far more developed in humans.)

On emotions, LeDoux starts by bemoaning the terminological mess that exists any time emotions are discussed.  He reserves the word “emotion” for conscious feelings, and resists its application to the lower level survival circuitry, which he sees as non-conscious.  He points out that a lot of published results which claim to show things such as fear in flies, are actually just showing survival circuit functionality.  He sees survival circuits as very ancient, going back to the earliest life forms, but emotions as relatively new, only existing in humans.

In LeDoux’s view, emotions, the conscious feelings, are cognitive constructions in the prefrontal cortex, predictions based on signals from the lower level survival circuitry, reinforced by interoceptive signals from the physiological changes that the lower level circuitry initiate: changes in blood pressure, heart rate, breathing, stomach muscle clenching, etc.

LeDoux’s views are similar to Lisa Feldmann Barrett’s constructive emotions theory, and contrast with views such as Jaak Panksepp, who saw consciously felt emotion in the lowest level survival circuits.  Barrett also sees emotions only existing in humans, although she makes allowances for animals to have affects, simpler more primal valenced feelings such as hunger, pain, etc.  I’m not sure what LeDoux’s position is on affects.  He doesn’t mention them in this book.

My views on all this is that I think LeDoux is too skeptical of animal consciousness.  It doesn’t seem like a human without language could pass his criteria.  However, as always, this may come down to which definition of “consciousness” we’re discussing.  Human level consciousness includes introspective self awareness and far wider ranging imagination, enabled by symbolic thought such as language, than exist in any other species.  If we set that as the minimum, then only humans are conscious, but many will see that as too stringent.  In particular, I think a case could be made that it’s far too stringent for sentience.

On emotion, I do think LeDoux is right that the lower level survival circuitry, the reflexes and fixed reaction patterns in subcortical regions, shouldn’t be thought of as feeling states.  This means we shouldn’t take defensive behaviors in simpler animals as evidence for fear, or aggressive behavior as evidence for anger.

On the other hand, I think he’s wrong that feeling states don’t come around until sophisticated deliberative processing.  It seems like any goal-directed instrumental behavior, such as selecting an action for a particular outcome, requires that there be some preference for that outcome, some valence, input from the lower level survival circuits to the higher level ones that decide whether to pursue a goal or avoid an outcome.

This might be far simpler than what humans feel, perhaps only meeting Barrett’s sense of an affect rather than what Barrett and LeDoux see as the full constructed emotion, but they should be felt states nonetheless.  By LeDoux’s own criteria, that would include any animal capable of instrumental behavior, including mammals and birds.  Admittedly, there’s no guarantee these felt states are conscious ones, but again, definitions.

Comparing LeDoux’s book to Feinberg and Mallatt’s, I’m struck by how much of the disagreement actually does come down to definitions.  The real differences, such as which species are capable of operant / instrumental learning, seem like they will eventually be resolvable empirically.  The differences on consciousness, may always be a matter of philosophical debate.

What do you think of LeDoux’s various stances?

Update 9-11-19: The statement above about LeDoux seeing instrumental learning only in mammals and birds isn’t right.  Please see the correction post.

Higher order theories of consciousness

I’ve posted on HOT (higher order thought theories of consciousness) before, but there’s a new paper out covering the basics of these types of theories.  Since first reading about HOT many months ago, the framework has been growing on me.  The paper is not too technical and I think would be accessible to most interested readers.

In general,  HOT is the idea that first order representations, the sensory images formed in early sensory areas of the brain, such as in the visual cortex, are not by themselves sufficient for being conscious of their contents.  We can hold and react to such representations unconsciously.

In order to be conscious of the representation, a second or higher order representation is needed, a representation about the first representation.  Inner awareness comes from these higher order representations accessing the first order representations.

Under most HOT theories, the higher order representations are in the prefrontal cortex.  This puts the PFC at the center of consciousness, a contentious view.  Although there is an openness to the possibility that some higher order representations might be in the parietal or temporal lobes.

I was a bit surprised that the paper described GWT (global workspace theory) as a first order theory.  GWT posits that for a perception to enter consciousness, it must be broadcast into the global workspace.  I would have thought that GWT was agnostic about how the information from the representation made it to the various cognitive modules.

I envisioned that GWT might provide a big picture view of what was happening, with HOT providing the details, and each of the cognitive modules holding their own versions of the representation, that is, their own representation of the initial one, tailored to their own needs.  But apparently GWT is more specific in its views about how the information is broadcast and received.  If so, that threatens to weaken GWT in my eyes, making it into a sort of data bus that I’m not sure matches the actual biology.

One of the things I do wonder about HOT is how accurate it is to call the higher order representations… representations.  It doesn’t seem controversial to say there is processing happening in other regions that make use of the information from the first order representations, or that being conscious of the representations requires processing outside of them.  But saying the information held by this other processing itself amounts to a representation might be an oversimplification.

Indeed, the division that many of these theories make between the representations and the processing that happens about the representations has long struck me as artificial.  I suspect there isn’t any clean line.  The representation probably gradually morphs through processing layers to being processing about associations of the representation and other related information.

Still, HOT strikes me as a productive way to think about inner experience.  We just have to remember that the higher order representations are not just mirror images of the first order ones, but are heavily tailored to the needs of the regions in which they exist.  Representations in the prefrontal cortex are likely integral components of action plans.

Interestingly, reading this paper, I realized that another theory I’ve long been fond of, Michael Graziano’s AST (attention schema theory), is itself a type of higher order theory, although I haven’t seen it described anywhere as one.  AST posits that the brain holds a representation, a schema of the process of attention, and that inner awareness results from this representation.

Graziano doesn’t see the schema as being in the prefrontal cortex, but in regions along the border between the parietal and temporal lobes, although the prefrontal cortex may be involved.  And again, various HOT theories seem open to the possibility of some higher order representations being outside of the prefrontal cortex.

HOT theories are almost certainly not the final word, if for no other reason than they differ on various points, and I suspect the devil is in those details, but the overall viewpoint strikes me as having promise.

Unless of course, I’m missing something?

Higher order theories of consciousness and metacognition

Some of you know, from various conversations, that over the last year or so I’ve flirted with the idea that consciousness is metacognition, although I’ve gradually backed away from it.  In humans, we typically define mental activity that we can introspect to be conscious and anything else to be unconscious.  But I’m swayed by the argument that mental activity accessible to introspection, but that we never get around to actually introspecting, is nevertheless conscious activity.

I had thought the idea of consciousness being metacognition was essentially what HOT (high order theories) of consciousness were all about, and so my backing away from the metacognition idea seemed to entail backing away from HOT.  However, a recent paper written to clear up common misconceptions about HOT points out that this is mistaken (page 5).

Just to review: metacognition is cognition about cognition, thinking about thinking, awareness of our own awareness, etc.  It’s essentially introspection and is necessary for introspective self awareness.

HOT, on the other hand, posits that there are two types of mental representations. There are simple representations about the external world, such as the neural pattern that forms in the visual cortex based on signals from the retina.  This would be a first order representation.  First order representations are often associated with early sensory processing regions and are not themselves sufficient for us to be conscious of them.

Then there are representations about these first order representations.  These are second order, or higher order representations, and are often associated with the prefrontal cortex, the executive center of the brain.

Under HOT, the contents of consciousness are these higher order representations.  The higher order representation is us being aware, conscious, of the first order representation.  (To be conscious of the higher order representation itself requires another higher order representation.)  Our sense of inner awareness comes from these representations of the representations.

Given that I’ve often pondered that qualia are the raw stuff of the communication from the perceiving parts of the brain (where the first order representations are) to the planning parts of the brain (where the second order representations are), HOT strikes me as very plausible.

Crucially, higher order representations, despite their name, are much more primal in nature than metacognition.  The paper does admit that they likely share some common functional areas, but metacognition is a more sophisticated and comprehensive activity.  It strikes me as very likely that metacognition is built with recursive higher order representations.

My only reservation with HOT, and I’m sure various specific versions handle this, is that not just any higher order representation is going to be conscious.  It will depend on where in the information flows that representation is formed.  That and the higher order representation shouldn’t be thought of a simple echo of the first order one, but as informational structures which have their own unique functionality.

Ultimately HOT will have to be judged by how well it matches observations, but a nice implication of it is that inner experiences aren’t ruled out for species that show little or no evidence for metacognition.