RM: Although PCT is about control of perception we seem to spend very little time (if any) discussing how perception works. If we want to build machines (robots) that can produce complex behavioral results we will, according to PCT, have to give those machines the ability to perceive the state of those results. From a PCT perspective, in other words, the central problem of robotics is designing perceptual, not output systems.
RM: Rather incredible progress in machine perception has apparently been made since I was studying perception in graduate school. For example, speech recognition systems were very poor back then; now I can talk into my phone and it will understand what I am saying with nearly great accuracy. And the speech perception system is speaker independent (as is the human speech perception system) and pretty good at understanding speech through different dialects. The same kind of progress has been made with handwriting recognition. I can now deposit checks by putting them in a slot and a machine will read the written amount of the check, again with astonishing accuracy.
This all came to my mind because my sister in law sent me a news article describing some recent work on machine perception. The article is actually about some fellows who proved that one approach to machine perception is equivalent to a process used in physics to describe systems without knowing the exact state of their components parts (whatever that means). The article is here:
RM: What’s interesting about the article (to me) is that it describes a perceptual learning machine that is based on a hierarchy of simulated neural networks, each level of the network computing a more complex aspect of the scene that is being perceived. Sound familiar?
RM: I doubt that the perceptual learning hierarchy referred to in the article is equivalent to the PCT hierarchy; certainly not in terms of the types of perceptual variables perceived. But I would like to know more about how machine perception works. Why,for example, is machine perception so much better than it was back in the 1970s and 80s. This may be mainly of interest to roboticists who will eventually have to design complex perceptual mechanisms in order to get their robots to do complex things. But if anyone knows anything about how some of these successful perceptual mechanisms work I would really like to hear about it. And if any robotists out there have built robots that control fairly complex perceptual variables it would be great if you could describe what those variables were and how they were computed. After all, understanding what (and how) organisms perceive is central to understanding what they are doing, according to PCT.
RM: I doubt that the perceptual learning hierarchy referred to in the article is equivalent to the PCT hierarchy; certainly not in terms of the types of perceptual variables perceived.
you’re probably wrong about that. what is the reason for your doubting the similarity between sufficiently evolved maths? the hierarchy is most likely the same (you probably didn’t think too hard about it/or you simply wouldn’t know enough about it to make the identification). thank you for giving me a link to the article. let me know when you’re done playing with your Lego mind storms and want to discuss these matters with a real engineer.
On Fri, Jan 16, 2015 at 2:15 PM, PHILIP JERAIR YERANOSIAN pyeranos@ucla.edu wrote:
RM: I doubt that the perceptual learning hierarchy referred to in the article is equivalent to the PCT hierarchy; certainly not in terms of the types of perceptual variables perceived.
PY: you’re probably wrong about that. what is the reason for your doubting the similarity between sufficiently evolved maths?
RM: The main reason is that the types of perceptions that are assumed to be computed at each level of the PCT hierarchy are based on Bill’s introspection about the types of perceptual variables he experienced and the apparent relationship between them. Actually, it would be interesting to see what types of perceptions are computed at each level of the Hinton hierarchy. I think it would be great if they turned out to correspond to the types in the PCT hierarchy.
PY: the hierarchy is most likely the same
RM: Again, that would be great. I think it would be possible to find out for sure by reading up on the Hinton hierarchy. I haven’t done it yet. It would be great if you could do this and report back on what you find.
Best
Rick
(you probably didn’t think too hard about it/or you simply wouldn’t know enough about it to make the identification). thank you for giving me a link to the article. let me know when you’re done playing with your Lego mind storms and want to discuss these matters with a real engineer.
–
Richard S. Marken, Ph.D.
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble
Bill’s hierarchy is basically correct. but it’s not coded precisely in mathematical terms so it’s very hard to compare with models built according to more exacting standards of rigor. It’ll take me a while longer to understand renormalization and the ising model but I’ve been studying their relation to pct for a while. These technques are used to describe many-system systems.
···
On Fri, Jan 16, 2015 at 2:15 PM, PHILIP JERAIR YERANOSIAN pyeranos@ucla.edu wrote:
RM: I doubt that the perceptual learning hierarchy referred to in the article is equivalent to the PCT hierarchy; certainly not in terms of the types of perceptual variables perceived.
PY: you’re probably wrong about that. what is the reason for your doubting the similarity between sufficiently evolved maths?
RM: The main reason is that the types of perceptions that are assumed to be computed at each level of the PCT hierarchy are based on Bill’s introspection about the types of perceptual variables he experienced and the apparent relationship between them. Actually, it would be interesting to see what types of perceptions are computed at each level of the Hinton hierarchy. I think it would be great if they turned out to correspond to the types in the PCT hierarchy.
PY: the hierarchy is most likely the same
RM: Again, that would be great. I think it would be possible to find out for sure by reading up on the Hinton hierarchy. I haven’t done it yet. It would be great if you could do this and report back on what you find.
Best
Rick
(you probably didn’t think too hard about it/or you simply wouldn’t know enough about it to make the identification). thank you for giving me a link to the article. let me know when you’re done playing with your Lego mind storms and want to discuss these matters with a real engineer.
–
Richard S. Marken, Ph.D.
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble
The parts most relevant to PCT seem to me to be between 9:25-16:00 and 17:25-20:05. The decomposition of images via a particular kind of unsupervised learning into a hierarchy of perceptions - first edges, then feature parts, then objects - seems tantalizingly close to the PCT hierarchy of intensities, sensations and
configurations (though feature parts and objects are probably closer to different aspects of configurations). The algorithm appears to apply equally well to speech and touch.
RM: This is really interesting Frank! Thanks. I'm not sure I understand how they came up with the "base set" of features; I presume it's similar to factor analysis with the base set of features being the "factors" that account for the "variance" in the pixel arrays of the different images.Â
I think the hierarchical levels are iterations of this "factor analysis" approach. So the next level up from "edges" are features (factors) that account for the variance of the factors in each image.Â
It's pretty clever stuff, though I'm not sure the levels of their hierarchy are really equivalent to those proposed in the PCT hierarchy. All the levels are really perceiving the same type of variable -- a configuration -- but computing those perceptions in a hierarchical manner. There are really not different _types_ of perceptual variables at each level. But it's still a very powerful pattern recognition system and probably a good model to use for a system that controls for producing particular configurations.Â
Â
BestÂ
Here is a link to a more technical presentation:
<http://research.microsoft.com/en-us/events/fs2013/andrew-ng_machinelearning.pdf>> http://research.microsoft.com/en-us/events/fs2013/andrew-ng_machinelearning.pdf
And here is a link to a tutorial:
<Bad title - Ufldl; UFLDL Tutorial - Ufldl
There is much in these that I donât understand, but it might give you some idea how the magic of human perception is being modeled.
Frank
From: "<mailto:csgnet@lists.illinois.edu>csgnet@lists.illinois.edu" <<mailto:csgnet@lists.illinois.edu>csgnet@lists.illinois.edu>
Reply-To: Richard Marken <<mailto:rsmarken@gmail.com>rsmarken@gmail.com>
Date: Friday, January 16, 2015 at 2:29 PM
To: "<mailto:csgnet@lists.illinois.edu>csgnet@lists.illinois.edu" <<mailto:csgnet@lists.illinois.edu>csgnet@lists.illinois.edu>
Cc: Richard S Marken <<mailto:richard.s.marken@aero.org>richard.s.marken@aero.org>
Subject: Putting the "P" back in PCT modeling
[From Rick Marken (2015.01.16.1230)]
RM: Although PCT is about control of perception we seem to spend very little time (if any) discussing how perception works. If we want to build machines (robots) that can produce complex behavioral results we will, according to PCT, have to give those machines the ability to perceive the state of those results. From a PCT perspective, in other words, the central problem of robotics is designing perceptual, not output systems.Â
RM: Rather incredible progress in machine perception has apparently been made since I was studying perception in graduate school. For example, speech recognition systems were very poor back then; now I can talk into my phone and it will understand what I am saying with nearly great accuracy. And the speech perception system is speaker independent (as is the human speech perception system) and pretty good at understanding speech through different dialects. The same kind of progress has been made with handwriting recognition. I can now deposit checks by putting them in a slot and a machine will read the written amount of the check, again with astonishing accuracy.Â
This all came to my mind because my sister in law sent me a news article describing some recent work on machine perception. The article is actually about some fellows who proved that one approach to machine perception is equivalent to a process used in physics to describe systems without knowing the exact state of their components parts (whatever that means). The article is here:
RM: What's interesting about the article (to me) is that it describes a perceptual learning machine that is based on a hierarchy of simulated neural networks, each level of the network computing a more complex aspect of the scene that is being perceived. Sound familiar?Â
RM: I doubt that the perceptual learning hierarchy referred to in the article is equivalent to the PCT hierarchy; certainly not in terms of the types of perceptual variables perceived. But I would like to know more about how machine perception works. Why,for example, is machine perception so much better than it was back in the 1970s and 80s. This may be mainly of interest to roboticists who will eventually have to design complex perceptual mechanisms in order to get their robots to do complex things. But if anyone knows anything about how some of these successful perceptual mechanisms work I would really like to hear about it. And if any robotists out there have built robots that control fairly complex perceptual variables it would be great if you could describe what those variables were and how they were computed. After all, understanding what (and how) organisms perceive is central to understanding what they are doing, according to PCT.Â
BestÂ
Rick
--
Richard S. Marken, Ph.D.
Author of  <Amazon.com Research on Purpose.Â
Now available from Amazon or Barnes & Noble
···
--
Richard S. Marken, Ph.D.
Author of  <Amazon.com Research on Purpose.Â
Now available from Amazon or Barnes & Noble