Wednesday, June 25, 2014

The IIT is falsifiable

Sadly I have come to a period of procrastination - five straight months of work, and here it is. It's understandable; in coming up with some final analyses for what is sure to be one of the best papers I ever put together, I've allowed the dimensionality of my data to explode, at the same time that I'm trying to rewrite the code that produces that data so that it can be run on a supercomputer, and at the same time that I'm trying to generate a poster on the work with an eye to current developments. So, the paper has, for this week at least, ground to a halt, and I'm sitting here staring at pages of code and thousands of .mat files and making spreadsheets to try to force some organization on the process.

Okay, deep breath.

I'm doing some highly integrative neuroscience - I'm applying complex information-theoretic measures to human neural data, in such a way that behavioral data produced by those humans can guide my interpretations. It's complicated.

The theory I'm working under is called the Integrated Information Theory (IIT) of consciousness. Consciousness is a fraught area of science and philosophy - everyone has an opinion, even if they think they don't - and the people who know their opinions are really, really excited about them. Some of those people are kind of weird. But the IIT, I think, is actually a rather mundane theory - it's very abstract on the one hand, but it doesn't, by its nature, make any big metaphysical claims about existence, or try to tie itself to evolution or quantum physics. Rather, the IIT aims to be a theory of the neural correlates of consciousness (NCCs) - why is it that some parts of the brain correspond to conscious experience, while others don't? At this level, it competes with other theories of NCC, most notably the Global Neuronal Workspace theory (GNW), although they take opposite approaches - the GNW looks first at the brain, at the NCCs themselves, and tries to explain the specific neural processes that we can see, objectively, are producing consciousness. The IIT, on the other hand, looks first at conscious experience, and, from the inside-out, tries to explain what properties consciousness has by proposing an algorithmic or mathematical expression of that experience.

I know the IIT better than the GNW, and I prefer the approach of the IIT, but I think that eventually these two theories will meet in the middle, and we'll have a real, general theory of consciousness. Might not be for a while, but it will happen.

What the GNW gives you is a description of the brain's mechanics, and post-hoc philosophizing about how certain types of connections in the brain distribute information throughout a central network, and that this distribution is the promotion of something to the status of consciousness. This is based on seeing, objectively, that these things are correlated, so the GNW is a theory of the NCCs at its very root. To some eyes, this makes it the more plausible competitor, since it's based in objective reality.

The IIT, being based in subjective experience (by extracting 'axioms' about consciousness - it's integrated, it's informative, it's bounded, it has structure) is different from the GNW especially in that it isn't specifically based on any principles of neuroscience. It's a mathematical theory, a theory of information transmission in networks. It is, of course, expressed with the obvious intention of being applied to facts of neuroscience, and it seems to do well in this respect. Put simply (I'm obviously not trying to explain the IIT here), if you point the IIT at a brain, it should give you back a prediction (you know what I mean) of whether or not that brain is conscious, which parts are doing the job, and how the doing is organized. This last part is one thing that makes the IIT so interesting - it gives a way of describing the internal structure of what it claims is conscious experience.

So, this is what the IIT is intended for - to predict how consciousness arises in a brain. In this it is indeed falsifiable. An attack I've seen on the IIT from several sources - most recently some of the duller commenters on Scott Aaronson's much better attack - is that it isn't falsifiable, but this is clearly not correct. Tononi (the theorist behind the IIT) gives one clear example of how to falsify the theory: take two substances, both of which disrupt neural activity, but only one of which degrades the capacity of the brain to integrate information (in the specific terms of the theory). You could equalize the substances in other ways - make it so that one impairs attention, or some other manipulation that under a given theory should destroy consciousness - and then find out whether or not your (human) subject has lost consciousness. The IIT should predict exactly which interventions disrupt consciousness and which do not. In fact, it should predict which parts of consciousness are degraded - is vision lost, or hearing, or etc. If the theory fails, then it's wrong. So, the IIT is falsifiable.

But when the critics say it's unfalsifiable, they aren't thinking about the object of the theory - human (or animal) consciousness. They're thinking about panpsychism, and thus missing the whole point. It's true - the IIT predicts that systems that aren't brains can possess consciousness, and that it can be completely alien to human consciousness - no perception or cognition, for example. To many people this is 1) totally incoherent (this is Aaronson's criticism) and 2) unfalsifiable and thus the mark of a bad theory. But that a theory generates unfalsifiable statements is never grounds for dismissing the theory - thsi is basic logic (cf. Popper). What matters it that, in the realm where it's meant to explain something, the theory does its job. If you have two theories that explain equally well, but one generates incoherent, untestable predictions, then you can proceed with parsimony, but you have to have the alternative first. Occam's razor can't be used when you have no competition (unless you'd rather have no theory at all - and this is an ideological point rather than a scientific approach to a problem).

And anyways, even the panpsychism problem (taking the meaning of panpsychism very loosely, since IIT doesn't predict that everything is conscious - just that things that aren't brains can be conscious) isn't definitely untestable - the theory predicts that different conscious things are, in principle at least, linkable. So, if IIT says that a certain not-a-brain system is conscious, then the IIT adherent can simply plug his brain into the object, and if the theory predicts that the adherent's consciousness is linked to the not-a-brain system, then firsthand knowledge is the proper test. That's science fiction stuff there (the film Pacific Rim's drift technology was an implementation of the idea), but the IIT says it's technically possible. So, the ultimate test of the IIT might turn out to be technology - if every IIT-based device fails to produce predicted effects on consciousness, then the theory will be abandoned.

That's all I have for now. June is almost over. One more month in Australia, then home. Back to the grind...

Tuesday, June 03, 2014

Sorry May

Okay, so this picture illustrates why I am not a Tegmarkian. Tegmark, if you don't know, is a clever cosmologist at MIT who's put forward (a book on) the thesis that mathematics is the ultimate reality, and that all mathematics is in fact a kind of reality - that there is a mathematical multiverse, which we know exists on account of the mathematics existing.

So I don't buy this. I'm diametrically opposed to this idea. Not opposed, really - I don't care too much, but I am opposed in that I believe the complete opposite. Mathematics - and physics as a subset of mathematics - is an artifact of the human mind, that's all it is. The fact that the world exists in some form is curious, although it seems incoherent to me that we can actually know anything about its true nature - but to suppose that its true nature is mathematics seems so backwards that I just wanted to write some things down.

I get where he's coming from. The world does exist, there is a reality, and it is somehow regular and consistent - it has properties that repeat or sustain, and why should it? Its continuities and discontinuities are all so numerically describable, and why should they be? And the most basic elements that we know to exist - photons, quarks, magnetic fields - seem to be perfectly and completely described as systems of numbers. And why should this be?

My mind seems to have taken the easy way out, because it just screams: but numbers and math are things that human minds *do*! They describe the world because the brain is a description machine, that's what it *does*! If the curious thing is that the description is so perfect and complete, then I have two responses - the space of possible descriptions that the mind can form is so vast, so impossibly vast, that it would be surprising if we could *not* find consistent systems of description for the world; and no description of the world is by any means *complete*.

The completion point is worth going on about. The scope and complexity of the natural world is impossible to comprehend. It's absolutely impossible to describe it all - and I'm saying this as a scientist with full faith in science as an endeavor for helping us to understand the world. We might choose some very narrow sliver of reality and subject it to intensive study, and then, there, we can describe it in such detail that we feel that it's okay to say we've basically got it all down. But that's it - those little, tiny, infinitesimally small splinters, and we think we have a complete description? What we have is a consistent system - mathematical physics - that can be used to describe anything we come across, but each description will be new, different, from what has been seen before.

So no description is complete. Okay, maybe that's a straw man, but I don't think so. Tegmark wants to claim that not only is physics a (potentially) complete description of our reality - or no, not a description, but *the thing itself* - but that realities we haven't yet encountered, i.e. realities *outside our reality* are contained within it. He likes the example of the discovery of Neptune. Astronomers had noted disturbances in the orbit of Uranus, and finally realized that there must be another planet even further out - they realized this mathematically, in such detail that they knew where to point their telescopes to find Neptune, and they did so, successfully.

Tegmark wants to use this example to imply that mathematics is a kind of tapestry containing all reality, and that by following it out from what was known, an *entire planet* was discovered, first in the mathematics, and only later by human senses. But this doesn't prove any kind of point about the reality of mathematics, and it's not even true, strictly, that Neptune was first discovered in a mathematical form. It was first discovered in the form of its gravitational influence, which affected Uranus. It's just that at first, astronomers didn't understand what they were seeing - they had to *do some math* in order to understand. But the data were all there - the measurements of Neptune in the flesh were there already, before Galle saw it with his own eyes (and others had seen it before, all the way back to Galileo, albeit not knowing what they were looking at).

The point here is that, really, new knowledge about the world can only come from new data about the world. Mathematics based on reality that has been observed - i.e. physics - can then tell you how to understand those data, but it is only that, a tool, an activity of the human observers. It doesn't exist outside of human endeavor. I am dead set in this opinion.

Anyways, so I basically had that conversation with myself last night on my walk home, and then I made that figure. It should be self-explanatory, but just in case: the biggest circle, the purple one, is the realm of all possible human thought. The circles within are not to any idea of scale, of course. There are many domains of human thought,and the next two that I've outlined are descriptions and axiomatic systems. Both of these I mean in the broadest sense you can imagine.  Physics falls within the realm of axiomatic systems of description, or it should (Hilbert's sixth problem). Within axiomatic systems you have consistent axiomatic systems, which should contain a correct physics, if it exists - i.e. if the Standard Model and General Relativity could be united. Taken as separate systems, I think that each of these theories alone counts as a consistent system, but together, so far, they do not.

Tegmark's reality is the domain of consistent axiomatic systems of description, of which our physics is (presumably) just a tiny part. Any other consistent system of physics would also fall in this domain, and Tegmark believes that each of these systems must also correspond to its own universe, just as our physics corresponds to ours. I think it's a fantastic idea, which I might illustrate by putting a big 'fantasy' circle somewhere in there, in between human thought and physics.