Connectionists: Brain-like computing fanfare and big data fanfare

Aurel A. Lazar aurel at ee.columbia.edu
Thu Jan 30 13:25:23 EST 2014


Brian,
the theoretical representation of the auditory and visual space with massively parallel neural
circuits (receptive fields + spiking neurons) in the spike domain is well understood. The mathematical
formalism uses the representation (encoding) of space-time signals in Hilbert space(s). You
can find pointers to the literature, decoding demonstrations of encoded video streams and,
Matlab and/or Python/PyCUDA open source code at: http://www.bionet.ee.columbia.edu/research/nce
Best,
Aurel
http://www.bionet.ee.columbia.edu




On Jan 29, 2014, at 1:02 PM, Brian J Mingus <brian.mingus at colorado.edu> wrote:

> Hi Włodek,
> 
> Suggesting that we use a Hilbert space instead of a standard vector space for semantic representation seems similar to the question of whether we should use spiking or rate-coded neurons. It would seem that if all of the information is encoded in the rate, that the Hilbert space representation could then be compressed into a simpler vector space representation. I did recently see a paper that suggested that different parts of the brain seem to use different codes, i.e., some convey information in spikes, and some convey information in the rate (I can't find the paper at this time). In that case, a Hilbert space representation might be simpler to the extent that it captures all of the information, either way.
> 
> I am not a physicist so I may not have a deep enough understanding of what a Hilbert space is - could you perhaps explain what new information it might be capable of representing over and above the vector space representation, which is essentially what is used in rate coded deep neural nets?
> 
> Thanks,
> 
> Brian
> 
> http://grey.colorado.edu/mingus
> 
> 
> 
> On Wed, Jan 29, 2014 at 3:26 AM, Włodzisław Duch <wduch at is.umk.pl> wrote:
> Dear all,
> 
>  
> 
> QM has yet to show some advantages over strong synchronization in classical models that unifies the activity of the whole network. There is another aspect to this discussion: we need to go beyond naïve interpretation of assigning functions to activity of single structures. We have to use a formalism similar to the quantum mechanical representation theory in Hilbert space, decomposing brain activations into combinations of other activations. In wrote a bit about it in sec. 2 of “Neurolinguistic Approach to Natural Language Processing”, Neural Networks 21(10), 1500-1510, 2008.
> 
>  
> 
> QM seems to be attractive because we do not understand how to make a transition between brain activations and subjective experience, described in some psychological spaces, outside and inside (3rd and 1st person) points of view. I have tried to explain it in a paper for APA,  Mind-Brain Relations, Geometric Perspective and Neurophenomenology, American Philosophical Association Newsletter 12(1), 1-7, 2012
> 
> 
> QM formalism of representation theory may be useful also for classical distributed computing systems.
> 
>  
> 
> Best regards, Włodek Duch
> 
> ____________________
> 
> Google W. Duch 
> 
>  
> 
>  
> 
>  
> 
> From: Connectionists [mailto:connectionists-bounces at mailman.srv.cs.cmu.edu] On Behalf Of Carson Chow
> Sent: Tuesday, January 28, 2014 10:01 PM
> To: connectionists at mailman.srv.cs.cmu.edu
> Subject: Re: Connectionists: Physics and Psychology (and the C-word)
> 
>  
> 
> Brian,
> 
> Quantum mechanics can be completely simulated on a classical computer so if quantum mechanics do matter for C then it must be a matter of computational efficiency and nothing more.  We also know that BQP (i.e. set of problems solved efficiently on a quantum computer) is bigger than BPP (set of problems solved effficiently on a classical computer) but not by much.  I'm not fully up to date on this but I think factoring and boson sampling or about the only two examples that are in BQP and not in BPP.  We also know that BPP is much smaller than NP, so if C does require QM then for some reason it sits in a small sliver of complexity space.
> 
> best,
> Carson
> 
> PS I do like your self-consistent test for confirming consciousness. I once proposed that we could just run Turing machines and see which ones asked why they exist as a test of C.  Kind of similar to your idea.
> 
> 
> On 1/28/14 3:09 PM, Brian J Mingus wrote:
> 
> Hi Richard, thanks for the feedback. 
> 
>  
> 
> > Yes, in general, having an outcome measure that correlates with C ... that is good, but only with a clear and unambigous meaning for C itself (which I don't think anyone has, so therefore it is, after all, of no value to look for outcome measures that correlate)
> 
>  
> 
> Actually, the outcome measure I described is independent of a clear and unambiguous meaning for C itself, and in an interesting way: the models, like us, essentially reinvent the entire literature, and have a conversation as we do, inventing almost all the same positions that we've invented (including the one in your paper). 
> 
>  
> 
> I will read your paper and see if it changes my position. At the present time, however, I can't imagine any information that would solve the so-called zombie problem. I'm not a big fan of integrative information theory - I don't think hydrogen atoms are conscious, and I don't think naive bayes trained on a large corpus and run in generative mode is conscious. Thus, if the model doesn't go through the same philosophical reasoning that we've collectively gone through with regards to subjective experience, then I'm going to wonder if its experience is anything like mine at all.
> 
>  
> 
> Touching back on QM, if we create a point neuron-based model that doesn't wax philosophical on consciousness, I'm going to wonder if we should add lower levels of analysis.
> 
>  
> 
> I will take a look at your paper, and see if it changes my view on this at all.
> 
>  
> 
> Cheers,
> 
>  
> 
> Brian Mingus
> 
>  
> 
> http://grey.colorado.edu/mingus
> 
>  
> 
>  
> 
> On Tue, Jan 28, 2014 at 12:05 PM, Richard Loosemore <rloosemore at susaro.com> wrote:
> 
> 
> 
> Brian,
> 
> Everything hinges on the definition of the concept ("consciousness") under consideration.
> 
> In the chapter I wrote in Wang & Goertzel's "Theoretical Foundations of Artificial General Intelligence" I pointed out (echoing Chalmers) that too much is said about C without a clear enough understanding of what is meant by it .... and then I went on to clarify what exactly could be meant by it, and thereby came to a resolution of the problem (with testable predictions).   So I think the answer to the question you pose below is that:
> 
> (a) Yes, in general, having an outcome measure that correlates with C ... that is good, but only with a clear and unambigous meaning for C itself (which I don't think anyone has, so therefore it is, after all, of no value to look for outcome measures that correlate), and 
> 
> (b) All three of the approaches you mention are sidelined and finessed by the approach I used in the abovementioned paper, where I clarify the definition by clarifying first why we have so much difficulty defining it.  In other words, there is a fourth way, and that is to explain it as ... well, I have to leave that dangling because there is too much subtlety to pack into an elevator pitch.  (The title is the best I can do:  " Human and Machine Consciousness as a Boundary Effect in the Concept Analysis Mechanism ").
> 
> Certainly though, the weakness of all quantum mechanics 'answers' is that they are stranded on the wrong side of the explanatory gap.
> 
> 
> Richard Loosemore
> 
> 
> Reference
> Loosemore, R.P.W. (2012).  Human and Machine Consciousness as a Boundary Effect in the Concept Analysis Mechanism.  In: P. Wang & B. Goertzel (Eds), Theoretical Foundations of Artifical General Intelligence.  Atlantis Press.
> http://richardloosemore.com/docs/2012a_Consciousness_rpwl.pdf
> 
> 
> 
> 
> On 1/28/14, 10:34 AM, Brian J Mingus wrote:
> 
> Hi Richard,
> 
>  
> 
> > I can tell you that the quantum story isn't nearly enough clear in the minds of physicists, yet, so how it can be applied to the C question is beyond me.  Frankly, it does NOT apply:  saying anything about observers and entanglement does not at any point touch the kind of statements that involve talk about qualia etc.
> 
>  
> 
> I'm not sure I see the argument you're trying to make here. If you have an outcome measure that you agree correlates with consciousness, then we have a framework for scientifically studying it. 
> 
>  
> 
> Here's my setup: If you create a society of models and do not expose them to a corpus containing consciousness philosophy and they then, in a reasonably short amount of time, independently rewrite it, they are almost certainly conscious. This design explicitly rules out a generative model that accidentally spits out consciousness philosophy.
> 
>  
> 
> Another approach is to accept that our brains are so similar that you and I are almost certainly both conscious, and to then perform experiments on each other and study our subjective reports.
> 
>  
> 
> Another approach is to perform experiments on your own brain and to write first person reports about your experience.
> 
>  
> 
> These three approaches each have tradeoffs, and each provide unique information. The first approach, in particular, might ultimately allow us to draw some of the strongest possible conclusions. For example, it allows for the scientific study of the extent to which quantum effects may or may not be relevant.
> 
>  
> 
> I'm very interested in hearing any counterarguments as to why this general approach won't work. If it can't work, then I would argue that perhaps we should not create full models of ourselves, but should instead focus on upgrading ourselves. From that perspective, getting this to work is extremely important, despite however futuristic it may seem.
> 
>  
> 
> > So let's let that sleeping dog lie.... (?).
> 
>  
> 
> Not gonna' happen. :)
> 
>  
> 
> Brian Mingus
> 
> http://grey.colorado.edu
> 
>  
> 
> On Tue, Jan 28, 2014 at 7:32 AM, Richard Loosemore <rloosemore at susaro.com> wrote:
> 
> On 1/27/14, 11:30 PM, Brian J Mingus wrote:
> 
> Consciousness is also such a bag of worms that we can't rule out that qualia owes its totally non-obvious and a priori unpredicted existence to concepts derived from quantum mechanics, such as nested observers, or entanglement.
> 
> As far as I know, my litmus test for a model is the only way to tell whether low-level quantum effects are required: if the model, which has not been exposed to a corpus containing consciousness philosophy, then goes on to independently recreate consciousness philosophy, despite the fact that it is composed of (for example) point neurons, then we can be sure that low-level quantum mechanical details are not important.
> 
> Note, however, that such a model might still rely on nested observers or entanglement. I'll let a quantum physicist chime in on that - although I will note that according to news articles I've read that we keep managing to entangle larger and larger objects - up to the size of molecules at this time, IIRC.
> 
> 
> Brian Mingus
> http://grey.colorado.edu/mingus
> 
> Speaking as someone is both a physicist and a cognitive scientist, AND someone who has written papers resolving that whole C-word issue, I can tell you that the quantum story isn't nearly enough clear in the minds of physicists, yet, so how it can be applied to the C question is beyond me.  Frankly, it does NOT apply:  saying anything about observers and entanglement does not at any point touch the kind of statements that involve talk about qualia etc.   So let's let that sleeping dog lie.... (?).
> 
> As for using the methods/standards of physics over here in cog sci ..... I think it best to listen to George Bernard Shaw on this one:  "Never do unto others as you would they do unto you:  their tastes may not be the same."
> 
> Our tastes (requirements/constraints/issues) are quite different, so what happens elsewhere cannot be directly, slavishly imported.
> 
> 
> Richard Loosemore
> 
> Wells College
> Aurora NY
> USA
> 
>  
> 
>  
> 
>  
> 
>  
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20140130/24a47c7f/attachment.html>


More information about the Connectionists mailing list