The people who work on predictive coding seem to think you don't need high-level "goals" or reinforcement learning. The system's only taks is to minimize prediction error, either by changing the prediction to match signals (perception) or by changing the signal to match predicitons (by taking action). (It is still a mystery to me how this may relate to an HTM system and I hope to understand this some day soon.) See, for example: "Reinforcement learning or active inference?" http://www.fil.ion.ucl.ac.uk/~karl/Reinforcement%20Learning%20or%20Active%20Inference.pdf
Thanks, - Jeff T From: Barry Maturkanich <[email protected]<mailto:[email protected]>> Reply-To: "NuPIC general mailing list." <[email protected]<mailto:[email protected]>> Date: Wednesday, August 21, 2013 9:35 PM To: "NuPIC general mailing list." <[email protected]<mailto:[email protected]>> Subject: Re: [nupic-dev] Mentioned presentation on action with CLA? Hi guys just thought I would chime in with some of my thoughts on this. I have thought about what it might mean to integrate motor commands to the CLA and one big question that I hit was how to decide which command to select? Let us say the CLA has learned predictions for many different motor sequences. So for example given my hand is in a particular position and the sensory input is in a given state, the CLA may have come to learn what will happen if I were to move my hand in any of 100 different ways. I may have very strong predictions for exactly what will happen in all those cases. So which motor command do I issue for my hand, if any? It doesn't matter that I know what will happen in all 100 cases, I still need *something* to decide, something to define the "why" I move one way and not another. This is why I think Michael is suggesting we need some sort of basal ganglia as it sure seems that is the place where this situation is resolved. Via reinforcement learning it seems to be figuring out which are the "good" states (as well as the "bad" states). So now given the 100 possible movements, the basal ganglia can act as a sort of gate through which only the movement associated with the largest believed reward is let through. Now we have our decision. It is constantly answering this question across every piece of cortex, which seems consistent with why all parts of cortex are connected to subcortical/basal ganglia via thalamus. This seems necessary since all areas of cortex need this "why" addressed at all times, else it will not know what to do (or not do) given many possibilities. This situation goes beyond motor out, I believe it is the same exact situation with simply deciding what to "think about" next. Again, all parts of cortex rely on the basal ganglia subcortical gate loop to resolve "why" think this or think that, or do this or do that when faced with many possible paths. I don't see how CLA with motor out could know which of many potential sequences to select if there were not some way of associating prediction states with desirability. This is great discussion by the way! Love this stuff. Barry On 8/20/13 5:19 PM, Michael Ferrier wrote: I agree with all of that Jeff, just have two points to add: - While all regions of the cortex seem to have some motor-related output, I think it could be misleading to say that all regions have some "motor command" output. Guillery and Sherman (e.g. http://ironzog.com/nupic/Guillery_Sherman_2002a.pdf) talk about layer 5 cells having output that branches either to the spinal cord, tectum, or pons. The tectum (superior colliculus in primates) is involved in orienting the eyes and head toward attention-grabbing stimuli and the pons relays information to the cerebellum, which is involved in the smooth coordination and timing of movement (but is not necessary to produce movement). The only cortical areas that send out "motor commands" to the spinal cord (and from there to muscles) are the motor areas (with a small percentage also coming from the parietal cortex and cingulate cortex). - While the cortex is in charge of recording and playing back patterns of movement, it's up to the basal ganglia (and its reinforcement learning mechanism) to determine which pattern to play back and when to play it back, based on the context of whatever else is active in the cortex. For this reason I would think that such a mechanism would be necessary for goal-directed movement. -Mike _____________ Michael Ferrier Department of Cognitive, Linguistic and Psychological Sciences, Brown University [email protected]<mailto:[email protected]> On Tue, Aug 20, 2013 at 5:19 PM, Jeff Hawkins <[email protected]<mailto:[email protected]>> wrote: Two neuro-scientists, Ray Guillery and Murray Sherman have pointed out that in every region of the neocortex they have looked, they find cells in layer 5 that project to muscles, the spinal cord, or other behavior related parts of the brain. For example in primary visual areas V1 and V2 there are layer 5 cells that project to the Superior Colliculus which generates saccades and other eye movements. I don’t believe they counted the basal ganglia as a “motor” destination. Sherman and Guillery have proposed that this is the normal state of affairs, that all areas of the cortex have a motor output. This is a beautiful idea and certainly mostly true. Sherman and Guillery have written extensively about these layer 5 cells. The axons from these cells split. One branch goes to the muscle or motor area and the other half goes to the next region up in the hierarchy. Thus all regions of the cortex have some motor output command, but that same command is passed up the hierarchy. The next region thus knows what behaviors are being generated. Layer 3 receives both sensory and motor input. Layer 3 is the primary feed forward layer. It is what I think of when thinking of the CLA. In the general case layer 3 is building a model of sensory data plus motor commands. Layer 5 is similar to layer 3 in many ways. I believe it is learning the same sequence of column activations and thus the same sequences. The unfolding patterns of layer 5 cells then associatively link to other motor areas and thus learn to control them. It is a bit hard to describe without images. Conventional wisdom says that the basal ganglia does not create behavior directly. It seems to be responsible for selecting between alternate motor plans stored in the cortex. I believe we can build a simple system consisting of one CLA representing layer 3 and another CLA representing layer 5. The Layer 5 CLA is driven by layer 3 and associatively links to some pre-existing motor generator. The system would learn to string together pre-existing behaviors in novel ways. I don’t know if we would need a basal ganglia equivalent. There are several unknowns but the basic idea seems sound. I have a talk that goes into this idea. We hope to record it and make it available. Jeff From: nupic [mailto:[email protected]<mailto:[email protected]>] On Behalf Of Michael Ferrier Sent: Tuesday, August 20, 2013 11:54 AM To: NuPIC general mailing list. Subject: Re: [nupic-dev] Mentioned presentation on action with CLA? The impression that I get from the neuroscience literature is that there are two basic types of learning in the brain. The first type could be called "model learning", it is what the cortex specializes in, and it's about learning hierarchical spatio-temporal models of input, from both external sensors and from other brain areas, representing the outside world, the body, and other internal states, and how they change over time. The second type is reinforcement learning, which uses built-in "reward" and "punishment" signals (such as pain or the taste of sugar) to learn what cortical patterns should be activated within a particular context of the activity of other cortical patterns, so as to maximize reward and minimize punishment. In the brain, reinforcement learning takes place in the basal ganglia, but uses input from many different areas of the cortex, and affects the activation of patterns within prefrontal and motor cortex to result in the control of attention, working memory and movement. For a more detailed discussion, see e.g. chapter 7 here: http://grey.colorado.edu/mediawiki/sites/CompCogNeuro/images/8/89/ccnbook_01_09_2012.pdf It's this dichotomy that I think the BECCA system is getting at, with their distinction between a "feature creator" and a "reinforcement learner". All cortical regions contribute in some way to motor output, if only by providing contextual information to the basal ganglia or to other subcortical structures involved in shaping motor output, such as the cerebellum or superior colliculus. But the final output to the spinal cord that actually produces movement comes mostly from the motor areas. CLA strikes me as being potentially a major advance in simulating the cortex and its spatio-temporal "model learning", but I think the addition of reinforcement learning will be necessary in order to approach the problems of action selection, attention, working memory and cognition in a brain-like way. -Mike _____________ Michael Ferrier Department of Cognitive, Linguistic and Psychological Sciences, Brown University [email protected]<mailto:[email protected]> On Tue, Aug 20, 2013 at 11:34 AM, Thompson, Jeff <[email protected]<mailto:[email protected]>> wrote: Hello SeH, While I appreciate your pointing out this open source project of which I was not aware, it seems to go against my question. I started paying attention to work on the CLA (again after many years) when I heard Jeff Hawkins speaking as he does below that "There are no pure “sensory” regions and no pure “motor” regions". It gave me hope that this work might avoid the pitfall of the classic "input-processing-output" loop of classic AI, which BECCA clearly seems to follow (see the attached diagram). We now know that there are just as many feedback connections going to back down to the "input" regions, and that action is not so different from perception (in that it uses similar machinery of prediction), and that "input" is intimately tied to the actions active during the input (instead of having "action" on the other side of world from "input", as in the BECCA diagram). I'm hopeful to see a diagram soon of many CLA modules for action and perception connected in a hierarchy which shows how action comes from similar prediction machinery as perception and how to avoid the pitfall of "input on one end, output on the other end." Thank you, - Jeff T ________________________________ From: nupic [[email protected]<mailto:[email protected]>] on behalf of SeH [[email protected]<mailto:[email protected]>] Sent: Monday, August 19, 2013 6:26 PM To: NuPIC general mailing list.; [email protected]<mailto:[email protected]> Subject: Re: [nupic-dev] Mentioned presentation on action with CLA? i imagined that something like OpenBECCA http://openbecca.org<http://openbecca.org/> could be integrated with NuPIC for a motor control system https://github.com/brohrer/becca<https://github.com/brohrer/becca/blob/master/doc/becca_0.4.5_users_guide.pdf> https://github.com/brohrer/becca/blob/master/doc/becca_0.4.5_users_guide.pdf from the opposite direction: part of BECCA's perceptual components may be enhanced (or replaced) with NuPIC https://github.com/brohrer/becca/blob/master/core/agent.py On Mon, Aug 19, 2013 at 8:25 PM, Thompson, Jeff <[email protected]<mailto:[email protected]>> wrote: Thank you for the quick reply. I'm in Los Angeles, so I hope someone does record your presentation at NASA. A similar question arose when I read "Thinking, predicting, and doing are all part of the same unfolding of sequences moving down the cortical hierarchy." (On Intelligence, p. 158.) I'm sure this is a FAQ, but do you have some writings or presentations about how a CLA would receive feedback signals coming down the hierarchy? Thank you, - Jeff From: Jeff Hawkins <[email protected]<mailto:[email protected]>> Reply-To: "NuPIC general mailing list." <[email protected]<mailto:[email protected]>> Date: Sunday, August 18, 2013 11:35 AM To: "'NuPIC general mailing list.'" <[email protected]<mailto:[email protected]>> Subject: Re: [nupic-dev] Mentioned presentation on action with CLA? Jeff, I wrote this presentation a couple years ago for a workshop on sensory motor integration. That workshop was held at the Santa Fe institute and I don’t believe it was recorded. The genesis of the workshop was a paper written by Murray Sherman and Raymond Guillery where they point out that every region of the neocortex (as far as they have looked) has cells in layer 5 that have a motor function. The big idea is that every region of neocortex does sensory inference and generates behavior. There are no pure “sensory” regions and no pure “motor” regions. It is one of those beautiful results that make you slap your head and say “of course!” I have always envisioned the CLA as modeling a section layer 3 in a region of the neocortex. Layer 3 is the primary input layer and is therefore doing inference on the input to that cortical region. Layer 5 is driven by layer 3 and has the cells that innervate muscles, or more often project to some sub-cortical area that generates behavior. I see how two CLAs, one for layer 3 and the other for layer 5 can work together to learn a sensory motor model of the world where today’s CLA is purely sensory. There is a lot I don’t understand but there is enough that I think we can make progress. I gave this presentation again earlier this year at Numenta. It wasn’t recorded. It looks like I might give it again this fall at NASA Ames here in Silicon Valley as there are a few roboticists there interested in it. I don’t mind recording it if someone could take care of the logistics. Jeff From: nupic [mailto:[email protected]] On Behalf Of Thompson, Jeff Sent: Saturday, August 17, 2013 4:57 PM To: NuPIC general mailing list. Subject: [nupic-dev] Mentioned presentation on action with CLA? Hello. In the introduction for the NuPIC Hackathon Kickoff, Jeff Hawkins talks briefly about the need for CLA integration with action. In response to a question, he says "We haven't done experiments with motor interaction. I have a presentation, I think about it." Is the presentation about motor interaction with CLA available? http://www.youtube.com/watch?feature=player_detailpage&v=yShNQvJEP6A&t=2188 Thank you, - Jeff Thompson _______________________________________________ nupic mailing list [email protected]<mailto:[email protected]> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected]<mailto:[email protected]> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected]<mailto:[email protected]> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected]<mailto:[email protected]>http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
