Thursday, August 10, 2017

how to make IIT (& other theories of consciousness) more respectable

last week i said i would post the talk i gave recently in Japan. so here it is.

i hope these points may be relevant and useful for other theories of C besides IIT too

When I learned that I was coming here to debate against IIT, I was really excited. By temperament I like debates, even when they get somewhat heated. There is much one can learn from debates if one doesn’t take scientific arguments personally.

But now that I better understand the purpose of this meeting, I feel I should try to be more constructive than usual. It is exciting to see that a community of consciousness researchers is starting to flourish here in Japan.

So what I will say here may sound like unsolicited advices at times. Of course, this may run the risk of sounding patronizing. But I hope to convince you that I have some unusual experiences that are worth sharing here.

Although many scientists work on consciousness in the US, I can say I am one of the few who had to try to get tenure in the relatively conservative Ivy League system while openly caring about nothing but consciousness. Many colleagues warned me that it was a dicey thing to do. To demonstrate credibility (i.e. avoid getting fired), just getting funding is not enough. Instead they said I must successfully break into the most conservative mainstream funding mechanisms (NIH/NSF). I was not good at it. Frankly, it was a very frustrating and disorienting experience.

But I’d like to think that through the process I’ve learnt a thing or two about how to convince our sternest critics that our work is not total nonsense. By having to put myself into their shoes, I have come to appreciate why they think so little of us sometimes.

Do they really think we are so bad?
To give an example, here is an actual quote from a senior colleague who responded to a post I shared on social media about IIT. This colleague is a highly influential theoretical neuroscientist with an H-index of around 40:

"... I haven't paid close attention but my impression of all the consciousness theories is that they're made up out of thin air, just 'here I describe some complicated stuff and now I'm going to pull out of my hat the magic claim that this set of complicated stuff equals consciousness’ with nothing more connecting the two than that. (And I thought Scott Aaronson's takedown of IIT was devastating.) Am I missing something?" (italics mine)

A publicity problem for all of us
It is probably fair to say, to the limited extent that IIT has gained acceptance, it has mostly appealed to researchers with a physics or mathematical background. Within mainstream neurobiology and psychology, it hasn’t done as well.

But I should point out that the colleague I quoted anonymously earlier is actually a physicist by training too. Among the mainstream computational theorists I’ve talked to, few are more positive. In fact many are even more critical. I will not quote them all here without permission, except it may be worth pointing out that the word “religious” often come up.

I feel this is unfortunate. Although I’m not exactly a fan of the theory, I find the basic ideas of IIT intriguing and intuitively interesting. In 2011 when David Rosenthal and I reviewed the literature, we included IIT as one of the major theories too. Right or wrong, the theory itself is certainly not unscientific.

But in the past few years, I think the general reception of IIT has gotten way more negative. As you can see from the quote, this may end up reflecting badly on the field as a whole - something for all of us to consider.

Instead of getting upset about how all this may be unfair and underserved, I think we should ask ourselves: what went wrong exactly?

Biomedical sciences should be data driven
When I meet someone who is a fan of IIT, I often like to ask them: what is your second favorite theory of consciousness? Not infrequently, the answer I get is: none. They say IIT is the only theory with this level of mathematical precision worthy of consideration.

But in general, the nature of biomedical sciences is such that precise mathematical formulation of theories is often not nearly as important as empirical evidence. Many influential hypotheses start out being qualitative. We started thinking about the heart as a pump well before we worked out the formal hydraulics; we know certain brain regions are important for language, not because we have precise mathematical definitions of how language works exactly.

This is not to say quantitative formulation isn’t important at all, but at most it is only important to the extent that it can be applied to actual data. This may be one reason why the mathematical complexity and abstruseness of IIT– way beyond what is currently testable directly w/o approximations - have not impressed many critics.

Also, in biomedical sciences, the conceptual appeal of a theory can only go so far. Many clever and appealing ideas turn out to be flat wrong. And often it takes years of accumulation of results of many studies from multiple labs to find out. In this sense, winning over our critics is particularly important. They need to be interested enough to test our theory too.

Some historical considerations
Regarding how appealing and intriguing ideas may turn out to be so wrong, perhaps there’s no better example than the downfall of the Freudian empire. To my mind, the modern science of consciousness is irrevocably colored by the devastating criticism made by Karl Popper that Freud was unfalsifiable. It is debatable whether the Freudian ideas themselves were actually always  unfalsifiable, and the Popperian notion of falsification has meanwhile lost traction within philosophy. But regardless, the accusation on Freudians’ misbehavior stuck. According to Hans Eysenck, this has set the entire field of psychology backwards for half a century. Within the field of consciousness we may think we are already over this, but our critics may not think so.

This is not to say I think IIT is in principle unfalsifiable. Not at all. But I think we would do well to always remember the importance of going out of our way to avoid this impression. As I will explain below, in this aspect perhaps we have not done as well as we could.

Current zeitgeist & the danger of over-promotion
Of relevance also is the fact that in psychology, as well as cognitive neuroscience in general, there is a perceived looming crisis of replicability. There have been many calls to make science more open & transparent, to be driven more by data shared among many rather than authoritative opinions by a few.

This issue is relevant to IIT, because as you know, the creator and a major proponent of the theory are two leading figures in the field, with great influence especially in the US - Giulio Tononi and Christof Koch. In turn, they have worked with Nobel laureates, Gerald Edelman and Francis Crick, respectively. Within the field we look up to these giant figures. And I personally owe Giulio and Christof enormous debt, both intellectual and otherwise.

But from the outside – i.e. from our critics’ point of view - one may not see the existence and importance of healthy disagreement as well as we do. Instead, what our leading figures say in public may well be taken as representative of the entire field, as reflecting the state of the art. This is especially the case when these statements are made in major media outlets such as the New York Times, alongside figures like the Dalai Lama. When it is said that IIT is the “only really promising fundamental theory of consciousness” (italics mine), with others echoing that it is a “leading” or “established” theory, our critics are naturally going to judge the theory against a very high bar.

But the question is, empirically speaking, is the theory ready for this level of public declaration? If not, we risk losing our credibility as a field in general.

A proposal for an entry requirement for new theories
So this may be the main substantive point I want to make today: Let us distinguish among three kinds of empirical predictions made by a theory.

The first kind are not really predictions, but rather hindsight explanations of known facts. The existence of some phenomena has been agreed upon on by virtually all theorists for decades, such as that the cerebellum may not be necessary for having perceptual experiences. Being able to explain these is good, but they don’t count as novel predictions, and they have only limited value in arbitrating among competing theories.

Then there are novel predictions on which competing theories are silent. For example, a cognitive / functionalist theory may not make predictions about the exact neuronal or molecular-level mechanisms of consciousness, so long as the relevant functions are somehow implemented. Whatever low-level mechanisms turn out to be true, such a theory won’t be falsified. Likewise, many theories are silent on the exact characteristics of the complexity of the substrate of consciousness, but only because they do not concern themselves with these issues. We can call such relevant predictions novel, but uncontested.

The third kind of prediction are both novel and contested. If confirmed, they provide the strongest support for a theory because they distinguish between theories. These predictions are contested because some major competing theories reject them. For example, many predictions made by the global workspace theory are in direct contradiction with the predictions made by Lamme’s local recurrence theory.

Global workspace theory predicts, for example, that -
  1. the prefrontal cortex is critically involved in generating the subjective experience in perception;
  2. that attention is necessary for conscious perception;
  3. that the phenomenology of conscious perception is sparse;
  4. and that consciousness is linked to specific higher cognitive functions such as task-switching and cognitive control.
Local recurrency theory predicts the exact opposite in all four cases. If a prediction goes one way, one theory will be correct and the other will be falsified. These predictions reflect the most meaningful and informative empirical questions one can ask about theories of consciousness.

What are the novel and contested predictions made by IIT? It seems that unlike the global workspace and local recurrency theories, most of the predictions IIT makes are of the first and second kind only.

In a field with so little data and so many theories already, perhaps it wouldn’t be a bad idea to set an entry requirement that any new theory must make novel and contested predictions (the 3rd kind). After all, global workspace and local recurrency have already met this bar. Just to be clear, I'm a fan of neither of these theories. But it will be doing the field a disservice to write off the achievements they have made in establishing this empirical standard. Before meeting this standard, perhaps no theory should be taken seriously - not to mention being promoted heavily in the popular media.

Some empirical challenges
So far I have talked mostly about the contexts in which IIT needs to be considered. Let’s talk about some actual scientific details too.

Early on, it was pointed out that complexity measures based (approximately) on IIT may be able to distinguish between different states of consciousness, such as being in a coma, an anesthetized state, or dreamless sleep (unconscious), versus being awake or in a dream (conscious). But these two kinds of states also differ greatly in terms of the level of ongoing cognition. It may be that IIT is a theory of consciousness in the sense that having active cognition is part of the phenomenon - as we use the term in everyday and clinical usage. But this is not what proponents of IIT say. They say it is promising because it is the only theory that can explain subjective perceptual experience per se. It starts from axioms concerning phenomenology, or so they say.

But it is well known that cognition or simple information processing can dissociate from subjective perceptual experience as in the case of blindsight. Do we have evidence that IIT is about subjective experience per se, not just sophisticated cognition?

A recent study showed that, regardless of whether one’s eyes are open or closed, so long as one is awake the complexity measures that are supposed to reflect consciousness in IIT terms are highly similar. I take it that when lying down at rest as an experimental subject, much on our conscious experience is visual. When eyes are closed, it is not that there is no visual experience whatsoever, but it is relatively dull and static. With eyes opened, visual experiences are rich and changing. How come the complexity measures don’t reflect such differences between eyes open vs close?

In another high-profile study published by proponents of IIT themselves, it was found that while driving and listening to the radio, the brain connectivity pattern as assessed by the same IIT-derived complexity measures seemed to break down into two separate networks. It is highly plausible that what broke down were streams of cognition rather than consciousness: the person never became two conscious entities with two separate experiential points of view as is supposed to happen in split-brain patients.

About split-brain patients, it is also said that IIT can give a principled account of their phenomenology: of a divided consciousness after the corpus callosum is severed. That is, the left hemisphere becomes only aware of the right visual field, and vice versa, and there’s no one unified consciousness of the whole visual field. But not all patients with severed corpus callosum show divided consciousness. A recent study reports such counter-examples. How can IIT explain cases like this?

Overall, there seem to be many studies which may on the surface look like they are in support of or compatible with IIT, but once we go into the details of the actual data they don’t work so well. For instance, in one study on impaired consciousness in vegetative and minimally conscious patients, the authors claimed that “disturbances in [connectivity between certain brain regions] have severe impact on information integration and are reflected in deficits in cognitive functioning probably leading to a total breakdown of consciousness” (italics mine; for the lead author's own take of the issue see this). But upon more careful reading, several of the supposedly important measures of network complexity properties are no different between the patient and healthy controls. Yet, in other studies comparing anesthetized state and normal wakefulness, these measures are taken  to account for the difference in consciousness, and as evidence in support of IIT.

My point here is not to say the empirical case for IIT is already over. But we should ask ourselves, if these cases don’t falsify IIT, what kind of cases would do the job? Is there ever going to be the possibility that IIT may be just empirically wrong? In general, I don’t often hear proponents of IIT bringing up these challenges & addressing them head on. As I mentioned earlier, given the historical burden we unfortunately bear, we would really do well to go out of our way to avoid the impression that we brush aside empirical challenges as inconveniences.

Conceptual appeal
Perhaps one may say: we can worry about these empirical matters as we go along, but from the outset, we should pay attention to the theory anyway because of its conceptual appeal.

But like I said, theoretical appeal in biomedical sciences in general should only go so far. Even such a clever man as Nobel laureate Linus Pauling once advocated that vitamin C could cure cancer, not because his mathematical models of molecules weren’t elegant enough, but because he too easily accepted sloppy empirical evidence in favor of his theoretical views. Given the precarious position our field is in, only solid empirical grounds can help us avoid the impression of sycophancy and ideological indulgence.

But perhaps the most important consideration is how do we assess conceptual appeal, given its highly subjective nature. If we restrict ourselves to talking among proponents of the theory, of course we will find much we like to hear within this echo chamber. But as Scott Aaronson pointed out, it may not be so intuitively appealing or plausible that IIT predicts that some simple sets of inactive logic gates are conscious. Many others find it just nonsensical to say that a smartphone is ‘somewhat’ conscious, regardless of what IIT predicts. According to the philosopher Adam Pautz, he couldn’t even make sense of what the theory means. Far from being a leading theory, to Pautz the theory “hasn’t even made it out of the gate”.

Overall, philosophers have worked on these kinds of conceptual issues for centuries. IIT simply has not won them over as a group. As such, unilateral claims that IIT is the only theory which can explain the problem of qualia do the theory little favor.

Concluding remarks
On balance, I think the empirical & conceptual limitations of IIT are there, but our problem regarding the strategy of promotion may be even more critical.

One might think we should just focus on the science, and ignore what others may think. In this field, none of us are strangers to this romantic feeling of going against the grain, which at times even becomes our source of inspiration. In response, let me end with this quote from Thomas Kuhn:

"As in political revolutions, so in paradigm choice [in science]—there is no standard higher than the assent of the relevant community... this issue of paradigm choice can never be unequivocally settled by logic and experiment alone."

I hope I do not sound like a politician today. But my point is, in science, these factors do matter. They matter because, let us be realistic: we are not going to solve consciousness right here today. We need the field to have a future. And unlike in physics, we are not in a position to be able to afford a “high risk high reward” kind of strategy right now. Jobs and funding in our field remain very hard to come by for junior people, in the US especially.

But beyond all that, this is also a matter of principle rather than mere strategy. Just as we embrace democracy even if we find it maddeningly inconvenient at times, earning the respect of our respected colleagues, i.e. subjecting ourselves to peer review, is part of what we signed on to do as academic scientists. No New York Times articles or TED talks can or should replace this time honored mechanism to the ground truth.

That is why I have not attempted to go into the details of the theory. Like many I have problems with some specific details regarding those so-called axioms (e.g. exclusion), but we can wait for these to be fixed in the next versions of IIT. Because a theory can be modified and improved, IIT may well become more appealing in the future. Let us see - if only we don’t inadvertently kill it off right now, along with the entire field, with the way we promote and evaluate the theory.


  1. Overcoming the empirical limitations and providing solid empirical evidence for IIT will be great move forward for the field.

  2. Excellent piece, Hakwan. Very diplomatic, well-argued, and compelling.

  3. Great article. There are two points of view.

    1. True value of IIT
    I understand there are severe criticisms to IIT. But I believe actual data for evidence also will be shown in the near future. IIT's notable point is showing consciousness as can be measured. Detailed (zombie?) seem to be shown only to follow western intuition.

    Whether there are contested shouldn't be important. Global workspace is based on homunculus idea, and much far from actual human brain.

    2. Reconsideration of Hard Problem
    It's curious to see Aaronson's idea here. However, as Chalmers' idea was imperfect, I think his pretty hard problem idea had been meaningless.

    Regarding Hard Problem, it is another problem to discuss. Considering 'unit qualia' (which is minimum ones) may induce a new idea.

  4. Hakwan, have you written this up into a paper yet?

    1. hi Tam, no. not really intending to. blogposts are just as well?