Maybe it's a little early, but I'm not the only one interested.
Uwe wrote:
we have to think about the hierarchies of regions and the according visualization. How is the output of one region delivered to the next? There of course must be changes made to the visualizer. Barry sure has elaborated something regarding region feed-forward activation (in earlier versions that was column activation being propgated).
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I believe we can learn how to make the architecture of multi-regions by looking the functioning of the regions of association.
In the human cortex, the sensory information (vision, hearing, touch, etc.) converge to associative regions (A1). In his book, Hawkins illustrates this well (2004):
“For example, your cortex has areas that receive input from both vision and touch. It is thanks to association regions that you are able to be aware that the sight of a fly crawling up your arm and the tickling sensation you feel there share the same cause. Most of these areas receive highly processed input from several senses”.
By this, I mean that when we create a hierarchy of regions, we have to ensure that when an input from R1 to reach A1, A1 will activate the same (learnt together) pattern of columns in R2 that are associated with the concept (object) predicted in A1 (depending of the activation strength). This is the case of touch sensory inputs representing a fly walking on our skin are "strong" enough to activate the columns of the regions that represent the visual image of a fly walking on the skin: https://sourceforge.net/p/openhtm/discussion/htm/thread/ea289de7/#
IT
/\
/ \
/ \
A1 ...
/ \
/ \
R1 R2
For this to occur properly both representations should have been correlated (learned together) in a "moment" in the cortex.
Is in this "moment" of learning that we could use a similar technique to the temporal pooler (ie instead of learning that was active at previous time, A1 should memorize what was active in another region however in the same time sequence).
In the above article and others, you will notice a term not addressed (in formal rules) in the CLA white paper: Bayesian inference. So I wrote a brief post about what are Bayesian networks and why HTM is a kind of:
I think when we consider multiple regions we need to get into the biology of how regions are really connected in animals. I recently went to a lecture in which large scale regional map of bird and human brains were presented. The regions where shown in a stylized circular diagram with regions presented at on the circumference. Many regions were connected across the "diameter" of the circle so that the connectivity between regions was more like a graph than a tree. I think the tree-like diagrams in Jeff Hawkin's book are actually not realistic (they are the weakest part of his argument, IMO), as real connections in animals simply don't look like that: there is no convergence upwards towards a single "high level" region. For the same region, I also take issue with Hawkin's analogy of an army with lower regions being the footsoldiers and the top region being the general. There is no evidence for this sort of hierarchy in any animal I have seen -- in fact the term hierarchy may be misleading in itself as it implies a tree structure.
Another thing the lecturer (Prof Murray Shanahan of Imperial College London) emphasised is the brain has a modular small world network arrangement. The word "modular" here indicates that the regions communicate with each other only through a small "hub" node (rather than each neuron in one region talking to each neuron in another region). I asked if the hub node was a single soma of a neuron but he said it was still a large number of neurons (though smaller than the number of neurons in a region).
So basically the question of multiple regions is complicated. I don't think Hawkin's or Numenta's publications shed much light on it -- and even neuroscientists don't know the whole story. It all very well to show a macroscopic map of a bird brain as described above, but each line in the map represents millions of axons bundled together. How this bundling occurs and how connections are made to individual neurons at either end of such a region-connecting bundle of axons remains, I think, a mystery.
Returning to the drawing I mentioned with the bird regions shown in a circle, many regions are connected across the centre of the circle. Prof Shanahan considers the connections in the centre to be vitally important and calls this a global workspace. Shanahan is a believer in Baars Global Workspace Theory which is worth investigating. Projects such as LIDA are attempts to implement this theory.
EDITED (by David): I simply hate the sourceforge forum sometimes. I had written a entire post and SF ignored it and still attached a image in a wrong place! John, only you can delete the attached images. :-(
You let me curious and decided research about this. :-) I discovered that both representations (graph and tree) can be valid, although that the tree is used only to represent small sets of cortical areas.
I found a lecture from Shimon Edelman (University of Cornell) which he also talks about "small world" cortical conections that you cited. And really, he initially presents a representation from cortex that doesn't remember a Tree at all (see Merker04-fig1.jpg). So your observation is valid.
However he took as sample a set of areas from a convoluted area and then explained in progressive steps how those areas can be organized in a simple hierarchy similar to a tree (actually he calls "pyramid-like" representation) (see Merker04-fig2.jpg).
So I believe we should assume that yes the cortex (as whole) is not a symetric tree, BUT still so we could face its specialized areas (in an isolated analysis) as a tree-like hierarchy. This way, when these same trees are connected each other, we have see the big picture of the cortex as a extremely complex "skein of cat", and not a big "redwood".
After someone mentioned Emergent, I took a closer look and found their wiki-book http://grey.colorado.edu/CompCogNeuro/index.php/CCNBook/Main very interesting. In it I've found many details about brain structure and here I'd like to post an image regarding activation flows between regions. It's about real brain and for feed-back connections just additional cortex layers are used, maybe this can help us in feed-back design (after we study Numenta's ideas of course)
Thanks for your comment. It's interesting that a graph can be described as a tree, I wasn't aware of that. I guess its possible provided you allow full data flow back down the tree so the top node acts as a facilitator of communication rather than just storing it. (I don't like the army General analogy!)
Still it might be better to think of the "graph" right from the beginning to avoid confusion. Of course it comes down to how important we think biological correctness is. When making an AGI, do we consider we are making an "animal" or not?
where he describes the brain's "connective core" in detail.
Shanahan believes in Baars Global Workspace theory which is a rival to HTM (but perhaps the two can coexist). Folks at Memphis University are implementing this in LIDA
Looks like LIDA is full of separate objects that each one is doing it's own thing and have it's own responsibility. how much code do you have to write in order to implement this? and how do you know the code you write is the correct one?
This theory about a "global workspace" might be right, but then it is a separate object from the neocortex, or the very top of the region hierarchy. however without a separate object that provides motivation and goal chasing, a prediction framework is only as good as giving global predictions about the world.
I think the place to learn is from biology experiments on mammals neocortexes, and another (suprising?) place to look for are the similarities between mammals and birds brains. looks like they might have something similar to neocortex with columnar structure.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Yes, its full of separate objects, in fact one of the developers describes it as a "kitchen sink". However, the latest issue of New Scientist (18 May 2013, dedicated to "consciousness") is absolutely full of Global Workspace theory. This theory is becoming hugely influential -- although that doesn't mean it's right!
The LIDA source (in Java) is available for free from Memphis Uni link I gave above. It's very hard to understand...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
By creating openHTM to inter-operate with other htms in real-time by standard interfaces we could see a thing like Baars Global Workspace emerge as htms are trained for specific tasks and combined to solve more and more abstract problems.
If openHTM standards are adopted I would hope to see the community of users/trainers expand exponentially. This might happen when openHtm regions have a common i/o format and easy way to hook input of a region to output of some other running htm region, and a internet registry (app store like?) of endpoints, or URL type endpoint. Regions could be running in the same machine or on another machine on the network or internet. As this network of regions grows, the higher layers will operate on more abstract meaning, and at a slower speed because changes from lower layers will take longer to propagate to higher regions. In this type of configuration we would see graph like structures that resemble a social network emerge.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
So far, all i can think of is predicting on the top region, then its just output to input output to input all the way down, then you see the result.
i figure you can tell the difference between a predicted output from a spacial output by simply naming your proximals as feedback possible or feedforward only.
that could give you some funny home video. :)
Last edit: Magnus Wootton 2013-05-17
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Maybe it's a little early, but I'm not the only one interested.
Uwe wrote:
I believe we can learn how to make the architecture of multi-regions by looking the functioning of the regions of association.
In the human cortex, the sensory information (vision, hearing, touch, etc.) converge to associative regions (A1). In his book, Hawkins illustrates this well (2004):
“For example, your cortex has areas that receive input from both vision and touch. It is thanks to association regions that you are able to be aware that the sight of a fly crawling up your arm and the tickling sensation you feel there share the same cause. Most of these areas receive highly processed input from several senses”.
By this, I mean that when we create a hierarchy of regions, we have to ensure that when an input from R1 to reach A1, A1 will activate the same (learnt together) pattern of columns in R2 that are associated with the concept (object) predicted in A1 (depending of the activation strength). This is the case of touch sensory inputs representing a fly walking on our skin are "strong" enough to activate the columns of the regions that represent the visual image of a fly walking on the skin:
https://sourceforge.net/p/openhtm/discussion/htm/thread/ea289de7/#
For this to occur properly both representations should have been correlated (learned together) in a "moment" in the cortex.
Is in this "moment" of learning that we could use a similar technique to the temporal pooler (ie instead of learning that was active at previous time, A1 should memorize what was active in another region however in the same time sequence).
Hawkins and Dileep also comment about this interaction of regions in this academic article:
http://scholar.google.com.br/scholar_url?hl=en&q=http://citeseerx.ist.psu.edu/viewdoc/download%3Fdoi%3D10.1.1.132.6744%26rep%3Drep1%26type%3Dpdf&sa=X&scisig=AAGBfm1NkNHBvWKf2C6-d3JZtlPqbGFlYQ&oi=scholarr&ei=SLdqUfqUE5Oa8wTMzYH4Dg&ved=0CCcQgAMoADAA
In the above article and others, you will notice a term not addressed (in formal rules) in the CLA white paper: Bayesian inference. So I wrote a brief post about what are Bayesian networks and why HTM is a kind of:
https://sourceforge.net/p/openhtm/discussion/neuroscience/thread/7d81f075/
Last edit: David Ragazzi 2013-04-14
I think when we consider multiple regions we need to get into the biology of how regions are really connected in animals. I recently went to a lecture in which large scale regional map of bird and human brains were presented. The regions where shown in a stylized circular diagram with regions presented at on the circumference. Many regions were connected across the "diameter" of the circle so that the connectivity between regions was more like a graph than a tree. I think the tree-like diagrams in Jeff Hawkin's book are actually not realistic (they are the weakest part of his argument, IMO), as real connections in animals simply don't look like that: there is no convergence upwards towards a single "high level" region. For the same region, I also take issue with Hawkin's analogy of an army with lower regions being the footsoldiers and the top region being the general. There is no evidence for this sort of hierarchy in any animal I have seen -- in fact the term hierarchy may be misleading in itself as it implies a tree structure.
Another thing the lecturer (Prof Murray Shanahan of Imperial College London) emphasised is the brain has a modular small world network arrangement. The word "modular" here indicates that the regions communicate with each other only through a small "hub" node (rather than each neuron in one region talking to each neuron in another region). I asked if the hub node was a single soma of a neuron but he said it was still a large number of neurons (though smaller than the number of neurons in a region).
So basically the question of multiple regions is complicated. I don't think Hawkin's or Numenta's publications shed much light on it -- and even neuroscientists don't know the whole story. It all very well to show a macroscopic map of a bird brain as described above, but each line in the map represents millions of axons bundled together. How this bundling occurs and how connections are made to individual neurons at either end of such a region-connecting bundle of axons remains, I think, a mystery.
Returning to the drawing I mentioned with the bird regions shown in a circle, many regions are connected across the centre of the circle. Prof Shanahan considers the connections in the centre to be vitally important and calls this a global workspace. Shanahan is a believer in Baars Global Workspace Theory which is worth investigating. Projects such as LIDA are attempts to implement this theory.
EDITED (by David): I simply hate the sourceforge forum sometimes. I had written a entire post and SF ignored it and still attached a image in a wrong place! John, only you can delete the attached images. :-(
Last edit: David Ragazzi 2013-05-14
Hi John,
You let me curious and decided research about this. :-) I discovered that both representations (graph and tree) can be valid, although that the tree is used only to represent small sets of cortical areas.
I found a lecture from Shimon Edelman (University of Cornell) which he also talks about "small world" cortical conections that you cited. And really, he initially presents a representation from cortex that doesn't remember a Tree at all (see Merker04-fig1.jpg). So your observation is valid.
However he took as sample a set of areas from a convoluted area and then explained in progressive steps how those areas can be organized in a simple hierarchy similar to a tree (actually he calls "pyramid-like" representation) (see Merker04-fig2.jpg).
So I believe we should assume that yes the cortex (as whole) is not a symetric tree, BUT still so we could face its specialized areas (in an isolated analysis) as a tree-like hierarchy. This way, when these same trees are connected each other, we have see the big picture of the cortex as a extremely complex "skein of cat", and not a big "redwood".
The Edelman lecture:
http://kybele.psych.cornell.edu/~edelman/Psych-4320/week-6.html
David
Last edit: David Ragazzi 2013-05-14
After someone mentioned Emergent, I took a closer look and found their wiki-book http://grey.colorado.edu/CompCogNeuro/index.php/CCNBook/Main very interesting. In it I've found many details about brain structure and here I'd like to post an image regarding activation flows between regions. It's about real brain and for feed-back connections just additional cortex layers are used, maybe this can help us in feed-back design (after we study Numenta's ideas of course)
Last edit: Nick 2013-05-15
Thanks for your comment. It's interesting that a graph can be described as a tree, I wasn't aware of that. I guess its possible provided you allow full data flow back down the tree so the top node acts as a facilitator of communication rather than just storing it. (I don't like the army General analogy!)
Still it might be better to think of the "graph" right from the beginning to avoid confusion. Of course it comes down to how important we think biological correctness is. When making an AGI, do we consider we are making an "animal" or not?
Here is Prof Shanahan's website.
http://www.doc.ic.ac.uk/~mpsha/
It's well worth looking through his papers, particularly this one.
http://www.doc.ic.ac.uk/~mpsha/ShanahanPTRS2012preprint.pdf
where he describes the brain's "connective core" in detail.
Shanahan believes in Baars Global Workspace theory which is a rival to HTM (but perhaps the two can coexist). Folks at Memphis University are implementing this in LIDA
http://ccrg.cs.memphis.edu/
I've downloaded the LIDA code but struggling to understand it! The approach is extremely complex.
Looks like LIDA is full of separate objects that each one is doing it's own thing and have it's own responsibility. how much code do you have to write in order to implement this? and how do you know the code you write is the correct one?
This theory about a "global workspace" might be right, but then it is a separate object from the neocortex, or the very top of the region hierarchy. however without a separate object that provides motivation and goal chasing, a prediction framework is only as good as giving global predictions about the world.
I think the place to learn is from biology experiments on mammals neocortexes, and another (suprising?) place to look for are the similarities between mammals and birds brains. looks like they might have something similar to neocortex with columnar structure.
@Itay
Yes, its full of separate objects, in fact one of the developers describes it as a "kitchen sink". However, the latest issue of New Scientist (18 May 2013, dedicated to "consciousness") is absolutely full of Global Workspace theory. This theory is becoming hugely influential -- although that doesn't mean it's right!
The LIDA source (in Java) is available for free from Memphis Uni link I gave above. It's very hard to understand...
By creating openHTM to inter-operate with other htms in real-time by standard interfaces we could see a thing like Baars Global Workspace emerge as htms are trained for specific tasks and combined to solve more and more abstract problems.
If openHTM standards are adopted I would hope to see the community of users/trainers expand exponentially. This might happen when openHtm regions have a common i/o format and easy way to hook input of a region to output of some other running htm region, and a internet registry (app store like?) of endpoints, or URL type endpoint. Regions could be running in the same machine or on another machine on the network or internet. As this network of regions grows, the higher layers will operate on more abstract meaning, and at a slower speed because changes from lower layers will take longer to propagate to higher regions. In this type of configuration we would see graph like structures that resemble a social network emerge.
So far, all i can think of is predicting on the top region, then its just output to input output to input all the way down, then you see the result.
i figure you can tell the difference between a predicted output from a spacial output by simply naming your proximals as feedback possible or feedforward only.
that could give you some funny home video. :)
Last edit: Magnus Wootton 2013-05-17