Showing posts with label Integrated Information Theory. Show all posts
Showing posts with label Integrated Information Theory. Show all posts

Thursday, October 19, 2017

Tononi Koch Test for Machine Consciousness

































In follow up to my previous post and before I saw Blade Runner 2049, I wanted to post a more modern take on the Turing Test based on a coherent theory of consciousness  by Tononi and Koch - both experts in the neuroscience of consciousness.  Their theory is the Integrated information Theory (IIT) of consciousness.  I have included the reference (1) and a graphic from their public access paper on the theory and there are also several very useful videos available to listen to the verbal descriptions of the theory.  I have been following consciousness research for at least the past 20 years including the two main listservs on this topic until they shut them down.  When a topic is so specialized, barring any breakthroughs the arguments become repetitive and a lot of time is spent bringing novices up to speed.  The videos fill a useful gap that these listservs previously addressed although I must admit  that I am always biased toward the written rather than the spoken word because it is a much more efficient information transfer for me.  The videos listed at the bottom of this page also serve another useful purpose.  The viewer is able to see how researchers in this area define consciousness and describe their theories.  I think that it is possible to notice that some of the definitions and descriptions are so vague as to have limited utility.

That is one of the reasons that I like the approach by Koch and Tonini.  I will also also say from the outset that I am not sure whether they view the theory as a joint venture or not.  As an example of what I mean looking at this specific search on consciousness finds that Tononi has been working in this area for at least 20 years.  A similar search on Koch goes back even 8 years earlier.  I don't know either of the authors but based on reading this paper it seems like a joint effort and that seems to come across in  the available videos of their presentations. (see addendum).

In the paper, that authors outline phenomenological definitions that are more exacting than any that I have seen in the past from other authors.  They are also neuroscience based and that makes a difference to me.  In various venues people often faintly praise but then lament psychiatry's emphasis on biology.  That is obviously not true or at least without reason and it also illustrates the lack of research that people do when it comes to critiquing psychiatry.  Psychiatrists have actively researched practically all forms of social, psychological, and biological etiologies of mental illness since the specialty was founded.  Any cursory review of a general psychiatric text illustrates that point.  So if a psychiatrist is focused on brain biology, it is certainly not without reason.  I previously posted a breakfast that I had with a mentor and after a long career as a psychiatrist he summed it up the way a lot of psychiatrists do: "It is all about the biology."  Critics take that to mean some kind of medical intervention.  They are certainly studied, but every other non-medical intervention has been studied as well.  It is common to read about non-medical interventions (psychotherapy, meditation, etc) altering the brain in some way.  In psychiatry that has been known within the field for at least 70 years.

There are two levels to study the work of Tononi and Koch.  The first is at the purely descriptive level.  That is the level that you will find in the first reference.  The second level is at the level of neuroscience and mathematical theory.  The authors have produced this work as well and reference it in this paper, but for the purpose of this post I am going to stay at the descriptive level and possibly post a more technical article on the advanced theory at a later date.  I will add that there are several competing theories of consciousness that I am not going to mention here.  I have studied several of them and think that they have less to offer than the Integrated information Theory (IIT) of consciousness.  I am admittedly a reductionist seeking to close the explanatory gap between brain biology and how conscious states are generated.  In some of the videos available online where there are panel discussions it is clear that the proponents of the other theories think that their own theories are correct and IIT is wrong. I have been down the rabbit hole with a few of those theories and don't want to take time to criticize them.  Feel free to look them up and form your own opinion.  For now I will focus on IIT.   

If you have never heard of Tononi, Koch, or IIT the first task is to read the paper.  I found it to be very clear in terms of definitions, postulates, and a clearly stated theory.  They point out that every experience will have an associate neural correlate of consciousness (NCC). There is currently an explanatory gap at the level of how conscious experiences are actually produced by the NCC.  They discuss the axioms necessary for a coherent phenomenology of consciousness.  From there they move on to the postulates.  Eventually they discuss how a conceptual structure that is maximally irreducible conceptual structure occurs in the brain.  These states are also known as quale.

They give a couple of examples about how conscious states occur within their theory.  They provide and example of how to calculate the quality and consciousness given a particular state containing elements (Figure 4).  They provide a clear example of the physical substrate of experience (complex), and a set of maximally irreducible cause-effect repertoires (concept), and a maximally irreducible "cause -effect structure in cause-effect space made of concepts..." or conceptual structure (quale)(p. 12).  The quantity of experience or consciousness is specified as Φmax.  The quality of experience is the form or shape of the conceptual structure. Distinct shapes occur with different experiences.

A more accessible example is discussed on page 9 and that is seeing Jennifer Aniston in a movie.  In that case, the complexes at the neuronal level affects the probability of past and future states. Consistent with neuroanatomy many specialized neurons are firing or not firing in the visual system that are associated with Jennifer Aniston as an invariant concept.  Other neurons are associated with other invariant concepts that allow for a fuller description in terms of appearance, age, etc.  All of the elements of the complex are intrinsic information and do not depend on visual inputs for example if dreaming or imagining the actress.

The authors also briefly review some of the experimental evidence that is consistent with the theory. They find that the theory is predictive in number of experimental paradigms. Transcranial magnetic stimulation (TMS) can be applied to to conscious individuals and unconscious (dreamless sleep, general anesthesia) individuals. In the conscious state there is a widespread pattern of activation and deactivation noted with high density EEG.  In the unconscious state cortical response is local or global and stereotypical - integration and information are lost.  A metric called the perturbation complexity index (PCI) a measure of the EEG compressibility from TMS stimulation can be used consciousness and it decreases in states that lack it.   

Tonini has been very explicit about the issue of machine consciousness - it doesn't exist no matter how sophisticated the machine is.  Any machine recognizing inputs that the human nervous system would recognize and producing identical outputs, even if that machine duplicates the structure and function of the human brain - is not conscious.  Tononi uses the consciousness science term zombie to characterize such machines.  By definition a zombie system is one that lacks consciousness and they are described as being subsystems in humans (2) when they are active outside the sphere of conscious recognition.

That brings us back to the ability to detect machines from humans.  If a machine is a perfect human zombie in terms of its input and output, we would not expect an empathy or Turing test to throw it off.   IIT theory acknowledges that what appears to be human input and output can be perfectly simulated.  The original Blade Runner protocol seems more than an empathy test. Specific questions about past memories illustrate an attempt determine if there is continuity between any current and past experiences, even though in the case of Rachael - the memories are false and implanted.

That being said IIT states there there is no Turing test for consciousness.  By now it does seem that fairly basic programs (like self learning neural nets) can replicate a narrowly defined human skill. In that case many people speculate that there is an intelligence or even human consciousness behind it.  On the other hand the perturbation complexity index (PCI) seems like a potentially useful test based on current results.



George Dawson, MD, DFAPA


References:

1: Tononi G, Koch C. Consciousness: here, there and everywhere?  Philos Trans R Soc Lond B Biol Sci. 2015 May 19;370(1668). pii: 20140167. doi: 10.1098/rstb.2014.0167. Review. PubMed PMID: 25823865; PubMed Central PMCID: PMC4387509.

2:  Koch C, Crick F. The zombie within. Nature. 2001 Jun 21;411(6840):893. PubMed
PMID: 11418835.




Addendum:

I read Christof Koch's book Consciousness - Confessions of a Romantic Reductionist a couple of months after this post.  In it he credits Tononi for Integrated Information Theory:

..."The theory of integrated information, developed by the neuroscientist and psychiatrist Giulio Tononi, starts with two basic axioms and proceeds to account for the phenomenal in the world." (p. 6)