[From Rupert Young (2015.05.25 11.30)]
I've been having a look at Deep Learning lately and have been struck
by some similarities with HPCT; simple units, hierarchical layers,
each unit in layer a function units in lower layer, weights between
units iteratively adjusted according to error. The main difference
being that error is not for control but a classification error
propagated down from the highest layer.
Deep Learning has mostly been applied to passive image recognition,
but here is an interesting video where it is applied to control,
Interestingly it gives the example of catching a baseball (2:38)
being solved by a parsimonious control rather than complex
prediction.
Also here is a video of approach to robotic learning fairly
compatible with PCT,
Anyone familiar with Deep Learning?
···
https://www.youtube.com/watch?v=EtMyH_–vnU
http://www.wirelessdesignmag.com/news/2015/05/robot-masters-new-skills-through-trial-and-error
–
**Regards,
Rupert
**
[From Rick Marken (2015.05.26.1140)]
···
Rupert Young (2015.05.25 11.30)–
RY: I've been having a look at Deep Learning lately and have been struck
by some similarities with HPCT; simple units, hierarchical layers,
each unit in layer a function units in lower layer, weights between
units iteratively adjusted according to error. The main difference
being that error is not for control but a classification error
propagated down from the highest layer.
Deep Learning has mostly been applied to passive image recognition,
but here is an interesting video where it is applied to control,
https://www.youtube.com/watch?v=EtMyH_–vnU
Interestingly it gives the example of catching a baseball (2:38)
being solved by a parsimonious control rather than complex
prediction.
RM: This was a very interesting talk but also very difficult. There are certainly some obvious similarities to reorganization learning. The “guided policy search” sounds a lot like E. coli reorganization where the “guidance” is provided by perception of error in the systems being reorganized. Maybe that’s what they mean by “optical control” where optimum control is minimum error value (cost, in their system?) resulting from the randomly selected “policies” (which I take to be analogous to the parameter settings of the control systems being learned (reorganized).
RM: The robots that result from this learning system are definitely control of input systems since they very nicely compensate for disturbances (most clearly demonstrated int he case of the walking robot). They call this “generalization” but it’s just what is expected from a well designed control system. I didn’t see what kind of initial structure is put into these robots, but then I really just skimmed the 1 hour talk and didn’t “deeply” understand a lot of what was being said anyway. But the results are very impressive; the towel folding robot was particularly interesting to me.
RM: It would be great if you could post something about how some of these robots work – in language that we (well, I) could understand. For example, since the robots being described seem to be perceptual control systems (based on their ability to keep variables like balance (location of center of gravity) and foldedness in reference states, protected from disturbances) where are the “reference specifications” for the states of these perceptions. Also, are they controlling different types of perceptions as the means of controlling higher level types of perceptions (eg. controlling pressure perceptions to control the visual image of a fold) or are they just controlling visual perceptions? I got the impression that it might be the latter.
RM: Oh, and the baseball catch stuff he refers to was done by our “competitors” – Dienes and McLoed. Actually, McLoed is at Oxford and I gave a PCT talk there in about 2005; he was not impressed. He doesn’t think of his model as a control model, even though it is. Ah well.
Best
Rick
–
**Regards,
Rupert
**
Also here is a video of approach to robotic learning fairly
compatible with PCT,
http://www.wirelessdesignmag.com/news/2015/05/robot-masters-new-skills-through-trial-and-error
Anyone familiar with Deep Learning?
–
Richard S. Marken
www.mindreadings.com
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble
Hi, this terminology seems to be current in neuroscience too as the free energy principle researchers talk about ‘policies’ and error reduction in a hierarchy…
Warren
···
Rupert Young (2015.05.25 11.30)–
RY: I've been having a look at Deep Learning lately and have been struck
by some similarities with HPCT; simple units, hierarchical layers,
each unit in layer a function units in lower layer, weights between
units iteratively adjusted according to error. The main difference
being that error is not for control but a classification error
propagated down from the highest layer.
Deep Learning has mostly been applied to passive image recognition,
but here is an interesting video where it is applied to control,
https://www.youtube.com/watch?v=EtMyH_–vnU
Interestingly it gives the example of catching a baseball (2:38)
being solved by a parsimonious control rather than complex
prediction.
RM: This was a very interesting talk but also very difficult. There are certainly some obvious similarities to reorganization learning. The “guided policy search” sounds a lot like E. coli reorganization where the “guidance” is provided by perception of error in the systems being reorganized. Maybe that’s what they mean by “optical control” where optimum control is minimum error value (cost, in their system?) resulting from the randomly selected “policies” (which I take to be analogous to the parameter settings of the control systems being learned (reorganized).
RM: The robots that result from this learning system are definitely control of input systems since they very nicely compensate for disturbances (most clearly demonstrated int he case of the walking robot). They call this “generalization” but it’s just what is expected from a well designed control system. I didn’t see what kind of initial structure is put into these robots, but then I really just skimmed the 1 hour talk and didn’t “deeply” understand a lot of what was being said anyway. But the results are very impressive; the towel folding robot was particularly interesting to me.
RM: It would be great if you could post something about how some of these robots work – in language that we (well, I) could understand. For example, since the robots being described seem to be perceptual control systems (based on their ability to keep variables like balance (location of center of gravity) and foldedness in reference states, protected from disturbances) where are the “reference specifications” for the states of these perceptions. Also, are they controlling different types of perceptions as the means of controlling higher level types of perceptions (eg. controlling pressure perceptions to control the visual image of a fold) or are they just controlling visual perceptions? I got
the impression that it might be the latter.
RM: Oh, and the baseball catch stuff he refers to was done by our “competitors” – Dienes and McLoed. Actually, McLoed is at Oxford and I gave a PCT talk there in about 2005; he was not impressed. He doesn’t think of his model as a control model, even though it is. Ah well.
Best
Rick
–
**Regards,
Rupert
**
Also here is a video of approach to robotic learning fairly
compatible with PCT,
http://www.wirelessdesignmag.com/news/2015/05/robot-masters-new-skills-through-trial-and-error
Anyone familiar with Deep Learning?
–
Richard S. Marken
www.mindreadings.com
Author of Doing Research on Purpose.
Now available from Amazon or Barnes & Noble