Well, I find this a bit disappointing. I was hoping that the contest between global workspace theory (GWT) and integrated information theory (IIT) would be announced sometime this year. Apparently, I’m going to have to wait awhile:
Pitts describes the intention of this competition as “to kill one or both theories,” but adds that while he is unsure that either will be definitively disproved, both theories have a good chance of being critically challenged. It’s expected to take three years for the experiments to be conducted and the data to be analyzed before a verdict is reached.
Three years. And of course there remains no guarantee the results will be decisive. Sigh.
I don’t particularly need the results to tell me which theory is more plausible. GWT is probably not the final theory, but it currently feels more grounded than IIT.
Still it would be nice to get insights on the back vs front of the brain matter. I’m expecting the answer to be the whole brain is usually involved, but that consciousness can get by in reduced form with only the posterior regions. (Assuming the necessary subcortical regions remains functional.) Having only those regions might be a sensory consciousness with no emotional feeling, a type of extreme akinetic mutism.
I suppose it’s conceivable consciousness could also get by with just the frontal regions, but it seems like it would be without any mental imagery (except for olfactory ones), blind feelings, which seems pretty desolate. On the other hand, it might amount to what the forebrain of fish and amphibians have, since their non-smell senses only go directly to their midbrain.
Oh well. I’m sure we’ll have plenty of other studies and papers to entertain us in the meantime!
31 thoughts on “The battle between integration and workspace will take a while”
I think I understand what IIT contends but I’m still hazy exactly what GWT contends.
For example, I just looked at a Baars paper reviewing evidence for the theory and he writes:
“This article reviews the scientific rediscovery of consciousness through the lens of global workspace theory and its central hypothesis, the notion that consciousness facilitates widespread
access between otherwise independent brain functions.”
To me that seems like just a description of what consciousness does. It doesn’t actually seem to explain it. And to me it doesn’t even seem to be that contradictory to IIT since it is describing essentially integration between brain functions.
Could you give us both theories as you understand in a nutshell without getting into where in the brain it all happens?
LikeLiked by 1 person
GWT contends that for a piece of information to enter consciousness, it must enter the workspace. To enter the workspace is to be broadcast throughout the brain (or more precisely the thalamo-cortical system). Every process is constantly trying to broadcast its content, so there is a competition, with a winner take all dynamic. When a winner manages to reach a threshold, all other competitors are inhibited, and the winner’s content is broadcast. What makes this broadcasted content conscious is its availability to all the unconscious processes, such as memory systems that might flag it for retention, emotional systems that might generate a reaction, introspection and language systems that might report it, etc.
IIT is often stated as contending that consciousness is integration of information. Put that simply, it doesn’t seem like it would contradict GWT. However, it includes a mathematical formula which purports to indicate the level of consciousness present. From what I understand, the mathematics seem to make it more of a theory about structural integration rather than information integration. And based on Christof Koch’s comments, it’s not really a theory about information processing, but about the brain’s “causal powers”.
This description may not be fair to IIT, but it’s the best I can manage. Many find the mathematics ambiguous and its predictions bizarre, such as Scott Aaronson noting that it’s trivial to construct a physical system with sky high phi (the calculated result of IIT’s calculations) which no one would regard as conscious. https://www.scottaaronson.com/blog/?p=1799
The reason I favor GWT is because 1) I understand it, or at least think I do, 2) the more I dig into it, including the neuroscience, the more sense it makes. GWT seems like a natural extension to cognitive neuroscience.
I can’t say I really understand IIT, and the more I try to understand it, the less sense it makes. I’m dependent to a large degree on Koch’s comments (I find Tononi’s parable oriented book unreadable). I don’t know; maybe it’s just been sold poorly.
Anyway, I’ve done posts on both of these theories:
Hope this helps.
LikeLiked by 2 people
Thanks. Maybe I understood it better than I thought.
I guess I’m still wondering “Where’s the beef?”
GWT seems to describe the circumstances surrounding something becoming conscious but doesn’t really specify the mechanism. Maybe that is just something we won’t ever know.
LikeLiked by 1 person
It’s actually a pretty simple architecture. Everyone, I think, quickly grasps the main idea. However, understanding why it works does take thought.
I think its central thesis is that consciousness is a community of unconscious specialist processes responding to the same information, with the output of each processes’ response vying to be, or be part of, the next set of information all the processes respond to.
Its strength is that it reduces consciousness to unconscious processes. However, this is in terms of access consciousness. It doesn’t really address phenomenal consciousness directly, except to consider it access consciousness from the inside. Michael Graziano’s attention schema added as one of the processes may help in explaining why we think there is something separate.
But if you’re convinced phenomenal consciousness is separate, then you won’t be satisfied.
James and Mike, I hope you don’t mind my taking this opportunity to put your last statements into the context of my current thinking. Yes, unitrackers. [reminder: a unitracker is a Mechanism, so, a set of neurons, which essentially tracks one thing (one pattern, one concept)]
“I think its central thesis is that consciousness is a community of unconscious specialist processes responding to the same information, with the output of each processes’ response vying to be, or be part of, the next set of information all the processes respond to.”
So this “community of unconscious specialist processes” pretty much lines up as a community of unitrackers. However, not all of these processes are necessarily vying to be in the next set of information. They may or may not. There can be action-based unitrackers that are essentially tracking opportunities to perform an action. When their criteria are met, they initiate the action (which may or may not be otherwise suppressed). For example, you probably have a unitracker for grasping and throwing a hand-sized object. When it’s target (an affordance for throwing, possibly including some plan or motivation for throwing) is recognized, it begins the reaching/grasping/throwing process. Another example might be a unitracker for generating a hormonal response (emotions, affects). There would probably be little use for these audience members to have their targets represented in the global workspace.
Now that I think of it, it may be the unitrackers which respond to the global workspace and do then compete to be represented in the global workspace are the “higher order thoughts” of HOT theories.
James Cross said:
“GWT seems to describe the circumstances surrounding something becoming conscious but doesn’t really specify the mechanism. Maybe that is just something we won’t ever know.”
The GWT proponents may not have proposed mechanisms, but there are already plausible mechanisms out there. I speak of Chris Eliasmith’s Semantic Pointer Architecture, which seems like a good candidate for a global workspace mechanism.
I like the idea of unitrackers. It conveys the idea that there neurons typically only get excited by certain things, such as lines, curves, shapes, combinations of shapes, faces, or a particular face. It also makes clear that in any particular mental event, most of the neurons aren’t participating.
“However, not all of these processes are necessarily vying to be in the next set of information. They may or may not. ”
Good point. At any one time, most of the composite unitrackers probably aren’t firing (or firing at higher than baseline), and so aren’t really vying to influence the workspace.
“Now that I think of it, it may be the unitrackers which respond to the global workspace and do then compete to be represented in the global workspace are the “higher order thoughts” of HOT theories.”
The responding processes may be high level composite ones, but they don’t have to be. Lower level processes could win also. (Although Dehaene points out that very low level processes often don’t have the connectivity.) But higher order ones probably have an advantage. A lower order unitracker’s best option might often be to excite a more summarized one and let it take on the burden of exciting the system.
Unitrackers seem implausible to me but maybe I don’t see how they relate to a term more commonly used “memory”. Memories seem to be distributed across the brain not in single neurons or sets of neurons. (For that matter, they may not even be stored in neurons.) Facial recognition seems to engage a broad network of neurons in multiple place in the brain not a unitracker for the face that is recognized.
The memory or perception of a particular thing is definitely spread out. When I talk about neurons getting excited by one thing, it’s important to understand that for a mom-face neuron to get excited, there first needed to be face-neurons excited, and for that there needed to be certain shape pattern neurons to get excited, which was dependent on certain shape, curve, color, and other related neurons first getting excited. The memory / perception of mom’s face is spread over those layers. But there is a neuron (or set of neurons) that uniquely recognize that this pattern is mom.
Memories are understood to be distributively stored in the varying strength of synapses. The forming of new memories requires new or strengthened synapses, forgetting requires weakening or disappearance of synapses. (There may also be memories of a fashion in the epigenetic state of some neurons, leading to particular combinations of protein formation.)
So, on faces in particular, a lot of neurons throughout many regions and layers are involved, but it’s neurons in the fusiform face area that are uniquely excited by faces. And when the mom-face neuron is excited, the dad-one, son-one, brother-one, etc, aren’t.
I don’t know enough about unitrackers in particular to defend them directly. I’ll have to leave that to James of S.
Yes. That is close to my understanding also but I’m not seeing how that relates to unitrackers. Just because the mom-face neuron is excited on seeing mom doesn’t necessarily mean that neuron is a unitracker for mom. In fact, it would make no sense that it would be because damage to that one single neuron would make recognition of mom impossible. That doesn’t correspond to our understanding of plasticity and redundancy in the brain. It would cause entire capabilities completely vanish with even minor damage in a small part of the brain.
To add to that example.
Let’s imagine a cluster of unitrackers for faces: mother, father, brother, old high school girl friend, an acquaintance from a college class who I did not know outside of the class. I hit my head getting out of my car. Now suddenly I can’t recognize my mother or my brother but I can recognize my father, the old high school girlfriend, and the acquaintance from college.
Would that make sense?
I think you’re overlooking the simple fact that, with plasticity, another neuron can step in, albeit perhaps after a brief period of confusion, particularly since many of the associations may be with aspects of the perception rather than with the final conclusion itself. And as I noted above, it’s probably not just one neuron, but a small cluster of them.
It’s also worth noting that lesions to the fusiform face area can leave us unable to recognize faces.
“lesions to the fusiform face area can leave us unable to recognize faces.”
That would make sense that ability to recognize faces in general might be impaired since we are dealing with a capability not a unitracker for a particular face.
James C, not sure why you think unitrackers are implausible. They’re just pattern recognizers, which is essentially what all the computer neural nets are. And they interact with each other as well as (sometimes) with the global workspace.
As for the mom unitracker, you probably have several. You probably have one for her as an individual person. But this unitracker will likely take input from a number of other unitrackers, including one for the idea of “mother”, and one for her full name, and one for her nickname, and one for her visual appearance (which can trigger regardless of angle of view), and perhaps one or more specific appearances from important memories of events, and one for her voice quality, etc. All of these will tend to trigger others, depending on current circumstances.
As for plasticity and memory, the hypothesis is that you are creating and losing unitrackers all the time. Yes, this is one kind of memory. If you lose one unitracker for a particular person’s face you could make a new one, unless you have lost the unitracker for “faces in general”, which apparently happens. See Oliver Sack’s “The Man Who Mistook His Wife for a Hat”, or just check the Wikipedia at https://en.m.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife_for_a_Hat.
Mike, you said:
“Good point. At any one time, most of the composite unitrackers probably aren’t firing (or firing at higher than baseline), and so aren’t really vying to influence the workspace.”
This suggests to me that you may have missed the point. There will be some unitrackers that have no input to the global workspace, so they never ever compete to be represented there. They simply have other functions.
I wanted to bring the unitracker thread up a few levels so there can be proper replies.
One of the things I find appealing about the unitracker idea is that it meshes with the anatomy of the neocortex. There is some reason to believe that the entire cortex is organized into units, also called columns. It seems plausible to me (even if it is an over-simplification) that one column could be one unitracker. Some of these unitrackers/columns will receive input directly from sensory modalities as well as from their unitracker neighbors. Much of the input from immediate neighbors will be inhibitory. Some might be from distant (prefrontal cortex?) unitrackers. Also, some (not all) will receive input from the global workspace.
So where is the global workspace? My oversimplified guess is that some (not all) unitrackers/columns will have exactly one axon going to the thalamus. That axon would have synapses on a large group of neurons in a region specific to the modality of that unitracker. Firing of that axon would tend to get those neurons firing in a pattern unique to that unitracker (constituting a “vector”, as described in the Semantic Pointer Architecture). Axons from these modality specific areas will go back to the unitrackers from that modality, but possibly also to other unitrackers, especially in the prefrontal cortex.
One of the puzzles I had about the relation between unitrackers and the global workspace involved a feedback loop. It would seem useful for the unitracker to activate its unique pattern in the workspace, but also that the unique pattern in the workspace could activate the unitracker. But this would seem seem to cause a feedback loop much like when an amplifier feeds back into a microphone. How could this work?
Enter the newly discovered XOR functioning neurons. An XOR means an exclusive or, which means it fires when it gets either input a, or input b, but not if it gets both at the same time. So what does that do? It could mean that the unitracker will fire if it gets input from its non-workspace sources, or if it gets input from the workspace, but not if it gets both. Thus, if its original input came from sensory, it will instigate its representation in the workspace, and then stop if it starts getting input back from the workspace (which means that it won the competition).
If unitracker is just a fancy name for a neural circuit that matches a pattern, then maybe I can live with it but the term is associated with Millikan and her philosophy which drags in with it a whole bunch of terminology that we then must translate to her framework. Why don’t you just call it a pattern matching neural circuit? BTW do we know that neural circuits like that exist for sure? What makes them match? Do small matches aggregate into bigger matches? Do all neural circuits match patterns? Or just some?
Regarding feedback and pattern matching. I do think you have something there. The question, however, is what happens when you are pattern matching to a pattern never before seen. In other words, what happens during learning. This seems to me a broad issue in neural circuit theories. A chicken and egg problem of sorts. If you can’t match with an existing circuit, you’re going to need a new circuit but there may, in fact, be no wiring for the new circuit. How do you fire the neurons together that are not wired together so they can match a pattern and then eventually wire together. This is where I think consciousness comes in with an EM field theory as the best explanation.
“Why don’t you just call it a pattern matching neural circuit?”
1. That’s 4 words instead of 1.
2. I don’t think “circuit” is an apt word for the mechanism.
3. Bringing in Millikan’s Framework is the plan. She’s already done the heavy lifting.
I’m pretty sure the answer to all those other questions is the work of neuroscience, and a lot has been worked out for the visual modality, although not down to the actual neuro mechanism. But what they have done seems compatible so far.
Regarding new unitrackers, I currently think pre-existing columns get recruited. If a new pattern shows up in the workspace and the “learning switch” is on, whichever column reacts most strongly will start getting reinforced, or alternatively some untrained one will be chosen. Thus, all those columns in the visual cortex of someone born blind will be available.
Sorry. I’m not sold on Millikan’s Framework at all. Honestly from what I can tell it is obfuscation. “Unitracker” seems to encapsulate unrelated processes like recognition and storage. “Same-tracking” seems like a term from A Clockwork Orange or Finnegan’s Wake.
Four words that describe what you are talking about are better than one word that lumps together unrelated processes apparently just so we manufacture a new term.
How do you difference-track the neurocepts and memunits in Millikan’s schema? Got you there! LOL
“seems to encapsulate unrelated processes like recognition and storage.”
Actually in the brain, the same neural circuitry that recognizes something is where that information is stored. The hippocampus facilitates synapse strengthening in those regions, but it actually happens in the circuitry where it’s used. It makes sense if you think about it. Why pay the metabolic cost twice? (Artificial neural networks work in a similar fashion, albeit virtually since the hardware they’re running on doesn’t.)
I didn’t mean to imply that I thought all unitrackers necessarily tried to feed into the global workspace. If I understand the concept correctly, many are lower parts of a hierarchy which wouldn’t produce the output for a particular overall process. But, just to make sure I’m understanding, what would you see as examples of ones that don’t?
On asking where the global workspace is, I think we need to be clear about a couple of things. First, the content of the workspace is not in any one location, but spread throughout the thalamo-cortical system. Any region with sufficient connectivity can win the competition and get its content into the workspace. But the content remains in the location in which it was generated.
We can talk about the location of the workspace in terms of where the major connection hubs are. As long as we understand that’s not where the content is, only where the signalling related to it passes through, we’re good. In that sense, it depends on who you talk to, but it’s usually held to be spread throughout the prefrontal cortex, precuneus in the parietal regions, along the cingulate gyrus, and possibly even parts of the temporal lobe. (And the thalamus, since I know that’s important to you.)
On single axons, I guess it depends on how narrow or broad the final unitracker that communicates with the workspace is, but I doubt it comes down to a single axon.
On feedback loops, actually the binding between source regions and receiving regions typically ends up being a recurrent connection, which reinforces the strength of signal from the source region and synchronizes the oscillations between regions. Of course, this can’t be a perfect loop since it would lead to a feedback like we see in audio systems. There has to be some bleed off so the signal can eventually die and be replaced by the next ascendant one. But recurrence is an important part of the binding process.
No idea if the XOR neurons figure into this.
Mike, I’m trying to figure out what you mean (or visualize) by a workspace. You say:
“First, the content of the workspace is not in any one location, but spread throughout the thalamo-cortical system.”
But then you say:
“[T]he content remains in the location in which it was generated.”
which seems contradictory.
What I mean is that visual content remains in the visual cortex, auditory content in the auditory cortex, touch sensation in the somatosensory cortex, language meaning in Wernicke’s area, etc. So the collection of the content the of the workspace is spread around, but any piece of content remains where it was computed.
When a touch sensation wins the competition, while the communication of that sensation is amplified by the global workspace, the source of that communication remains the somatosensory cortex, where it was initially computed. After winning, the region is able to bind with many other regions throughout the cortex. (The binding happening with the recurrent connections I mentioned before.)
Hope that helps.
Mike, it’s a start, but I want to know about possible mechanisms. When you say the source of a communication stays in the modal area in the cortex, that does not conflict with what I say. But you say “When a touch sensation wins the competition, while the communication of that sensation is amplified by the global workspace”, how could that work? And what do you mean by amplified? If the workspace is spread out thru the cortex, etc., then does the unitracker communicate to all of those neurons? Are the neurons of the workspace dedicated to being the workspace and nothing else?
“how could that work? ”
Region A computes content and sends out signals, both to other regions it’s directly connected to and to workspace hubs. The pathways into the hubs do two things, attempt to excite other regions, and attempt to inhibit other signals trying to do the same thing. Other destination regions, after receiving the signal from the originating region, may themselves get excited and send out their own signals compatible with the signal from the original. If enough regions do that, they form an effective coalition, reinforcing each other and inhibiting the other incompatible signals. If a threshold is reached, the workspace massively inhibits all the other signals (resulting in the P3 wave or “conscious ignition”) and clears the way for the winning signal(s). This gives the coalition regions clear access to bind to destination regions throughout the cortex.
“And what do you mean by amplified?”
Amplified was a bad word. Sorry. Just trying to use short hand for what I described above.
“If the workspace is spread out thru the cortex, etc., then does the unitracker communicate to all of those neurons?”
If we’re equating a unitracker to one of the processes, then yes, it forms recurrent connections with other processes throughout the cortex. Of course, not every process necessarily binds to ever other process, but based on some of the studies I’ve seen, the broadcasting is pretty pervasive, with, for example, visual processing regions being affected by motor action regions, etc.
“Are the neurons of the workspace dedicated to being the workspace and nothing else?”
The answer here may vary between variants of GWT. Dehaene’s global neuronal workspace posits that there are specialty regions dedicated to computing their set of tasks, and global communication regions dedicated to workspace dynamics.
The beginning of this paper (first two sections) gives a quick overview of the currently leading variant of GWT.
Click to access Dehaene_GlobalNeuronalWorkspace.pdf
[back to the top]
So now I’m having trouble with using the terminology of workspace and broadcast. My understanding of a workspace was a centralized location where representations could be made by agents/regions/unitrackers, and then others could access those representations. But the architecture you and the GWT folks suggest seems to be spread diffusely in and among said agents/regions/unitrackers. This workspace doesn’t so much contain a representation of the current contents of Consciousness, but instead coordinates among the various regions to determine who gets to “display/effuse/shout” unhindered, if briefly. It’s then each winning region which “broadcasts” independently. When I think of broadcasting, I think of one or a small number of central units sending information that can vary widely, from music to a sportscast. The model I see the GWT folks proposing is more like a bunch of HAM radio users, each of which sends exactly one message, “me”.
I don’t see the workspace, and I don’t see the broadcasting.
And now for nits.
1. Don’t worry about “amplified”, it’s all over that paper you linked.
2. You talk about processes “binding”. Not sure what you mean there. I know how to bind concepts in a semantic pointer architecture, but that architecture doesn’t seem to exist in the diffuse workspace described by GWT. So how does “binding” work?
3. I found section 6 of that paper, describing the anatomy of GWT the most useful.
To me, all of the empirical evidence described is compatible with a more centralized workspace. The global activity seen, including the “ignition”, would simply be the sequellae from a finalized representation in the workspace, wherever it is. I could even see moving that workspace up into the cortex, as GWT does, except that would introduce certain problems. Specifically, if the workspace as I see it was diffused throughout the cortex, damage to any significant part of the cortex would have disastrous effects on the ability of the workspace to represent a variety of information.
So I see two options: centralize the workspace, or stop calling it a workspace. 🙂
I can see the terminology angst. GWT originally started out as a psychological model based on an AI research concept called a “blackboard.” Which is to say, in its original incarnation, I suspect some central location where things stored probably was envisioned. I think it gradually morphed into the more distributed conception that’s out there today. I noted to Peter Carruthers on his post today about GWT that a better name might be “broadcast competition theory”, although it sounds like you have issues with “broadcast” too.
On binding, it’s basically the joining of related concepts. So if I see a pie, that might evoke a number or related things, such as memories of past pies, whether I like the flavor I’m seeing, etc. These might arise in various regions. The binding ties them together. The originating region forms circuits with the receiving regions. All the long range circuits in the cortex are reciprocal, so it leads to signals going back to the originating region. We end up with the two regions exchanging signals, reinforcing various aspects of whatever patterns they’re firing and sychronizing their activity. These recurrent connections are the bindings. Baars discusses it in this paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3664777/
“Specifically, if the workspace as I see it was diffused throughout the cortex, damage to any significant part of the cortex would have disastrous effects on the ability of the workspace to represent a variety of information.”
This assumes that the workspace has a brittle architecture, but the evidence seems to be that it is more resilient, able to continue functioning in a reduced capacity despite large sections missing. But damage to significant parts does have effects on its ability to convey a variety of information.
I agree on the workspace terminology, but I don’t think GWT is about to centralize (the data doesn’t point that way) and the “workspace” name appears to be too ingrained. We just have to understand that “workspace” refers to the entire cortical-thalamic system, although confusingly some researchers do use it to refer to the hub regions.
So Mike, have you seen this paper? (https://www.biorxiv.org/content/10.1101/2020.01.15.908400v1) “Distinguishing the neural correlates of perceptual awareness and post-perceptual processing”
From the abstract:
“Overall, we found a robust P3b in the report condition, but no P3b whatsoever in the no-report condition.”
I’m pretty sure the “P3b [signal]” is essentially the “global ignition” signal that the the GWT folks use to justify the global broadcast happening throughout the cortex. This paper suggests that the signal is instead associated with preparing to report an experience (the report condition), and the signal doesn’t happen when the experience happens under identical circumstances except that the subject is not required to report it (the no-report condition).
I haven’t read the report, so I could be wrong about the above, but the twitterati chatter suggests I might be right. Also, just one report. Bottom line: it surprises me none.
I was actually thinking about posting on it tomorrow. I did read it. It’s important to note that the authors are not claiming this disproves GWT, or HOT, or any other cognitive theory. Strictly speaking, it doesn’t even disprove global neuronal workspace theory (Dehaene’s specific version of GWT), although it does knock out a major piece of evidence for that specific variant. But it also raises an interesting possibility I’m investigating. Maybe more tomorrow.
Never said it disproved global workspace theory. I think said Theory is largely correct, and this paper would have no effect on my version, given my hypothesis of the location of said workspace. 🙂
Looking forward to what you come up with.
The post is up now.
One thing to consider about your hypothesis is whether it remains a global workspace theory? I’m an old programmer, so I understand how the “global” part gets interpreted in the sense of global variables as opposed to private or local variables. (It’s pretty much what I did when first hearing about the theory.) But I think theory means that the workspace itself is globally distributed.
In that sense, your variant may be more centralized workspace theory.