Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
I've been following this thread pretty much since the beginning. I hope I didn't miss anything subtle. You'll let me know if I have, I'm sure. ;=) It appears the need for temporal dependencies or different levels of reasoning has been conflated with the terms forward-chaining (FWC) and backward-chaining (BWC), which are typically used to describe different rule base evaluation algorithms used by expert systems. The terms “forward-chaining” and “backward-chaining” when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate “if/then” rules in a rule base (RB). In the FWC algorithm, the “if” part is evaluated and, if TRUE, the “then” part is added to the FWC engine's output. In the BWC algorithm, the “then” part is evaluated and, if TRUE, the “if” part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. To help remove any mystery that may still surround these concepts, here is an FWC algorithm in pseudo-code (WARNING: I'm glossing over quite a few details here – I'll be happy to answer questions on list or off): 0. set loop index to 0 1. got next rule? no: goto 5 2. is rule FIRED? yes: goto 1 3. is key equal to rule's antecedent? yes: add consequent to output, mark rule as FIRED, output is new key, goto 0 4. goto 1 5. more input data? yes: input data is new key, goto 0 6. done. To turn this into a BWC algorithm, we need only modify Step #3 to read as follows: 3. is key equal to rule's consequent? yes: add antecedent to output, mark rule as FIRED, output is new key, goto 0 If you need to represent temporal dependencies in FWC/BWC systems, you have to express them using rules. For example, if washer-a MUST be placed on bolt-b before nut-c can be screwed on, the rule base might look something like this: 1. if installed(washer-x) then install(nut-z) 2. if installed(bolt-y) then install(washer-x) 3. if notInstalled(bolt-y) then install(bolt-y) In this case, rule #1 won't get fired until rule #2 fires (nut-z can't get installed until washer-x has been installed). Rule #2 won't get fired until rule #3 has fired (washer-x can't get installed until bolt-y has been installed). NUT-Z! (Sorry, couldn't help it.) To kick things off, we pass in “bolt-y” as the initial key. This triggers rule #3, which will trigger rule #2, which will trigger rule #1. These temporal dependencies result in the following assembly sequence: install bolt-y, then install washer-x, and, finally, install nut-z. A similar thing can be done to implement rule hierarchies. 1. if levelIs(0) and installed(washer-x) then install(nut-z) 2. if levelIs(0) and installed(nut-z) goLevel(1) 3. if levelIs(1) and notInstalled(gadget-xx) then install(gadget-xx) 4. if levelIs(0) and installed(bolt-y) then install(washer-x) 5. if levelIs(0) and notInstalled(bolt-y) then install(bolt-y) Here rule #2 won't fire until rule #1 has fired. Rule #1 won't fire unless rule #4 has fired. Rule #4 won't fire until rule #5 has fired. And, finally, Rule #3 won't fire until Rule #2 has fired. So, level 0 could represent the reasoning required before level 1 rules (rule #3 here) will be of any use. (That's not the case here, of course, just stretching my humble example as far as I can.) Note, again, that the temporal and level references in the rules are NOT used by the BWC. They probably will be used by the part of the program that does something with the BWC's output (the install(), goLevel(), etc. functions). And, again, the results should be completely unaffected by the order in which the RB rules are evaluated or fired. I hope this helps. Cheers, Brad Richard Loosemore wrote: Mike Tintner wrote: A tangential comment here. Looking at this and other related threads I can't help thinking: jeez, here are you guys still endlessly arguing about the simplest of syllogisms, seemingly unable to progress beyond them. (Don't you ever have that feeling?) My impression is that the fault lies with logic itself - as soon as you start to apply logic to the real world, even only tangentially with talk of forward and backward or temporal considerations, you fall into a quagmire of ambiguity, and no one is really sure what they are talking about. Even the simplest if p then q logical proposition is actually infinitely ambiguous. No? (Is there a Godel's Theorem of logic?) Well, now you have me in a cleft stick, methinks. I *hate* logic as a way to understand cognition, because I think it is a derivative process within a high-functional AGI system, not a foundation process that sits underneath everything else. But, on the other hand, I do understand how it
Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
On Tue, Jul 15, 2008 at 8:01 AM, Brad Paulsen [EMAIL PROTECTED] wrote: The terms forward-chaining and backward-chaining when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate if/then rules in a rule base (RB). In the FWC algorithm, the if part is evaluated and, if TRUE, the then part is added to the FWC engine's output. In the BWC algorithm, the then part is evaluated and, if TRUE, the if part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. Curiously, the definition put by Abram Demski is the only one I've been aware of until yesterday (I believe it's the one used among theorem proving people). Let's see what googling says on forward chaining: 1. (Wikipedia) 2. http://www.amzi.com/ExpertSystemsInProlog/05forward.htm A large number of expert systems require the use of forward chaining, or data driven inference. [...] Data driven expert systems are different from the goal driven, or backward chaining systems seen in the previous chapters. The goal driven approach is practical when there are a reasonable number of possible final answers, as in the case of a diagnostic or identification system. The system methodically tries to prove or disprove each possible answer, gathering the needed information as it goes. The data driven approach is practical when combinatorial explosion creates a seemingly infinite number of possible right answers, such as possible configurations of a machine. 3. http://ai.eecs.umich.edu/cogarch0/common/prop/chain.html Forward-chaining implies that upon assertion of new knowledge, all relevant inductive and deductive rules are fired exhaustively, effectively making all knowledge about the current state explicit within the state. Forward chaining may be regarded as progress from a known state (the original knowledge) towards a goal state(s). Backward-chaining by an architecture means that no rules are fired upon assertion of new knowledge. When an unknown predicate about a known piece of knowledge is detected in an operator's condition list, all rules relevant to the knowledge in question are fired until the question is answered or until quiescence. Thus, backward chaining systems normally work from a goal state back to the original state. 4. http://www.ontotext.com/inference/reasoning_strategies.html * Forward-chaining: to start from the known facts and to perform the inference in an inductive fashion. This kind of reasoning can have diverse objectives, for instance: to compute the inferred closure; to answer a particular query; to infer a particular sort of knowledge (e.g. the class taxonomy); etc. * Backward-chaining: to start from a particular fact or from a query and by means of using deductive reasoning to try to verify that fact or to obtain all possible results of the query. Typically, the reasoner decomposes the fact into simpler facts that can be found in the knowledge base or transforms it into alternative facts that can be proven applying further recursive transformations. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
RE: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Ed Porter said: You imply you have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimensional and presumably large semantic space. Unfortunately I was unable to understand from your description how you claimed to have accomplished this. - I never implied that I have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimension and presumably large semantic space. I clearly stated: I have often talked about the use of multi-level complex methods and I see some similarity to the ideas that they discussed to my ideas. -and, The complex groupings of objects that I have in mind would have been derived using different methods of analysis and combination and when a group of them is called from an input analysis their use should tend to narrow the objects that might be expected given the detection by the feature detectors. Although I haven't expressed myself very clearly, this is very similar to what Riesenhuber and Poggio were suggesting that their methods would be capable of. So, yes,I think some similar methods can be used in NLP. I clearly used the expression in mind just to avoid the kind of misunderstanding that you made. I never made the exaggerated claim that I had accomplished it. The difference between having an idea in mind and having claimed to have accomplished a goal, which the majority of participants in the group would acknowledge is elusive, should be obvious and easy to understand. I am not claiming that I have a method that would work in all semantic space. I would be happy to claim that I do have a theory which I believe should show some limited extensibility in semantic space that goes beyond other current theories. However, I will not know for sure until I test it and right now that looks like it would be years off. I would be happy to continue the dialog if it can be conducted in a less confrontational and more genial manner than it has been during the past week. Jim Bromer Jim, In the Riesenhuber and Poggio paper the binding that were handled implicitly involved spatial relationships, such as an observed roughly horizontal line substantially touching an observed roughly vertical line at their respective ends, even though their might be other horizontal and vertical lines not having this relationship in the input pixel space. It achieves such implicit bindings by having enough separate models to be able to detect, by direct mapping, such a touching relationship between a horizontal and vertical lines at each of many different locations in the visual input space. But the Poggio paper deals with a relatively small number of relationships in a relatively small (160x160) low dimensional (2d) space using 23 million models. You imply you have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimensional and presumably large semantic space. Unfortunately I was unable to understand from your description how you claimed to have accomplished this. Could you please clarify you description with regard to this point. Ed Porter -Original Message- From: Jim Bromer [mailto:[EMAIL PROTECTED] Sent: Monday, July 14, 2008 1:38 PM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? I started reading a Riesenhuber and Poggio paper and there are some similarities to ideas that I have considered although my ideas were explicitly developed about computer programs that would use symbolic information and are not neural theories. It is interesting that Risesnhuber and Poggio argued that the binding problem seems to be a problem for only some models of object recognition. In other words, it seems that they are claiming that the problem disappears with their model of neural cognition! The study of feature detectors in cats eyes is old news and I did incorporate that information into the development of my own theories. I have often talked about the use of multi-level complex methods and I see some similarity to the ideas that they discussed to my ideas. In my model an input would be scanned for different features using different kinds of analysis on the input. So then a configuration of simple features would be derived from the scan and these could be associated with a number of complex groups of objects that have been previously associated with the features. Because the complex groups of objects are complexes (in the general sense), and would be learned by previous experience, they are not insipidly modeled on one standard model. These complex objects are complex in that they are not all cut from one standard. The older implementations that used operations that were taken from set theory on groups were set on object models that were very old-world and were not derived from learning. For example they were non-experiential. (I
Re: Location of goal/purpose was Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/14 Terren Suydam [EMAIL PROTECTED]: Will, --- On Fri, 7/11/08, William Pearson [EMAIL PROTECTED] wrote: Purpose and goal are not intrinsic to systems. I agree this is true with designed systems. And I would also say of evolved systems. My fingers purpose could equally well be said to be for picking ticks out of the hair of my kin or for touch typing. E.g. why do I keep my fingernails short, so that they do not impede my typing. The purpose of gut bacteria is to help me digest my food. The purpose of part of my brain is to do differentiation of functions, because I have . The designed system is ultimately an extension of the designer's mind, wherein lies the purpose. Oddly enough that is what I want the system to be. Rather an extension of my brain. Of course, as you note, the system in question can serve multiple purposes, each of which lies in the mind of some other observer. The same is true of your system, even though its behavior may evolve. Your button is what tethers its purpose to your mind. On the other hand, we can create simulations in which purpose is truly emergent. To support emergence our design must support large-scale, (global) interactions of locally specified entities. Conway's Game of Life is an example of such a system - what is its purpose? To provide an interesting system for researchers to research cellular automata? ;) I think I can see your point, It has no practical purpose as such. Just a research purpose. It certainly wasn't specified. And neither am I specifying the purpose of mine! I'm quite happy to hook up the button to something I press when I feel like it. I could decide the purpose of the system was to learn and be good at backgammon one day, in which case my presses would reflect that, or I could decide the purpose of the system was to search the web. If you want to think of a good analogy for how emergent I want the system to be. Imagine someone came along to one of your life simulations and interfered with the simulation to give some more food to some of the entities that he liked the look of. This wouldn't be anything so crude as to specify the fitness or artificial breeding, but it would tilt the scales in the favour of entities that he liked all else being equal. Would this invalidate the whole simulation because he interfered and bought some of his purpose into it? If so, I don't see why. The simplest answer is probably that it has none. But what if our design of the local level was a little more interesting, such that at the global level, we would eventually see self-sustaining entities that reproduced, competed for resources, evolved, etc, and became more complex over a large number of iterations? Then the system itself still wouldn't have a practical purpose. For a system Y to have a purpose, you have to have be able to say part X is like it is for Y to perform its function. Internal state corresponding to the entities might be said to have purpose, but not the system as a whole. Whether that's possible is another matter, but assuming for the moment it was, the purpose of that system could be defined in roughly the same way as trying to define the purpose of life itself. We have to be careful here. What meaning of the word life are you using? 1) The biosphere + evolution 2) And individuals exsistance. The first has no purpose. You can never look at the biosphere and figure out what bits are for what in the grander scheme of things, or ask yourself what mutations are likely to be thrown up to better achieve its goal. That we have some self-regulation on the Gaian scale is purely anthropic, biospheres without it would likely have driven themselves to a state not able to support lives. An individual entity has a purpose, though. So to that extent the purposeless can create the purposeful. So unless you believe that life was designed by God (in which case the purpose of life would lie in the mind of God), the purpose of the system is indeed intrinsic to the system itself. I think I would still say it didn't have a purpose. If I get your meaning right. Will --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Brad Paulsen wrote: I've been following this thread pretty much since the beginning. I hope I didn't miss anything subtle. You'll let me know if I have, I'm sure. ;=) It appears the need for temporal dependencies or different levels of reasoning has been conflated with the terms forward-chaining (FWC) and backward-chaining (BWC), which are typically used to describe different rule base evaluation algorithms used by expert systems. The terms “forward-chaining” and “backward-chaining” when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate “if/then” rules in a rule base (RB). In the FWC algorithm, the “if” part is evaluated and, if TRUE, the “then” part is added to the FWC engine's output. In the BWC algorithm, the “then” part is evaluated and, if TRUE, the “if” part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. Whooaa there. Something not right here. Backward chaining is about starting with a goal statement that you would like to prove, but at the beginning it is just a hypothesis. In BWC you go about proving the statement by trying to find facts that might support it. You would not start from the statement and then add knowledge to your knowledgebase that is consistent with it. So for example, if your goal is to prove that Socrates is mortal, then your above desciption of BWC would cause the following to occur 1) Does any rule allow us to conclude that x is/is not mortal? 2) Answer: yes, the following rules allow us to do that: If x is a plant, then x is mortal If x is a rock, then x is not mortal If x is a robot, then x is not mortal If x lives in a post-singularity era, then x is not mortal If x is a slug, then x is mortal If x is a japanese beetle, then x is mortal If x is a side of beef, then x is mortal If x is a screwdriver, then x is not mortal If x is a god, then x is not mortal If x is a living creature, then x is mortal If x is a goat, then x is mortal If x is a parrot in a Dead Parrot Sketch, then x is mortal 3) Ask the knowledge base if Socrates is a plant, if Socrates is a rock, etc etc . working through the above list. 3) [According to your version of BWC, if I understand you aright] Okay, if we cannot find any facts in the KB that say that Socrates is known to be one of these things, then add the first of these to the KB: Socrates is a plant [This is the bit that I question: we don't do the opposite of forward chaining at this step]. 4) Now repeat to find all rules that allow us to conclude that x is a plant. For this set of ... then x is a plant rules, go back and repeat the loop from step 2 onwards. Then if this does not work, Well, you can imagine the rest of the story: keep iterating until you can prove or disprove that Socrates is mortal. I cannot seem to reconcile this with your statement above that backward chaining simply involves the opposite of forward chaining, namely adding antecedents to the KB and working backwards. To help remove any mystery that may still surround these concepts, here is an FWC algorithm in pseudo-code (WARNING: I'm glossing over quite a few details here – I'll be happy to answer questions on list or off): 0. set loop index to 0 1. got next rule? no: goto 5 2. is rule FIRED? yes: goto 1 3. is key equal to rule's antecedent? yes: add consequent to output, mark rule as FIRED, output is new key, goto 0 4. goto 1 5. more input data? yes: input data is new key, goto 0 6. done. To turn this into a BWC algorithm, we need only modify Step #3 to read as follows: 3. is key equal to rule's consequent? yes: add antecedent to output, mark rule as FIRED, output is new key, goto 0 If you need to represent temporal dependencies in FWC/BWC systems, you have to express them using rules. For example, if washer-a MUST be placed on bolt-b before nut-c can be screwed on, the rule base might look something like this: 1. if installed(washer-x) then install(nut-z) 2. if installed(bolt-y) then install(washer-x) 3. if notInstalled(bolt-y) then install(bolt-y) In this case, rule #1 won't get fired until rule #2 fires (nut-z can't get installed until washer-x has been installed). Rule #2 won't get fired until rule #3 has fired (washer-x can't get installed until bolt-y has been installed). NUT-Z! (Sorry, couldn't help it.) To kick things off, we pass in “bolt-y” as the initial key. This triggers rule #3, which will trigger rule #2, which will trigger rule #1. These temporal dependencies result in the following assembly sequence: install bolt-y, then install washer-x, and, finally, install nut-z. A similar thing can be done to
RE: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Lukasz, Your post below was great. Your clippings from Google confirm much of the understanding that Abram Demski was helping me reach yesterday. In one of his posts Abram was discussing my prior statement that top-down activation could be either forward or backward chaining. He said If the network is passing down an expectation based on other data, informing the lower network of what to expect, then this is forward chaining. But if the signal is not an expectation, but more like a query pay attention to data that might conform/contradict this hypothesis, and notify me ASAP then it is backwards chaining. And it seems realistic that it can be both of these. I am interpreting this quoted statement as implying the purpose of backward chaining is to search for forward chaining paths that either confirm or contradict a pattern of interest or that provide a path or plan to a desired goal. In this view the backward part of backward chaining provides no changes in probability, only changes in attention, and it is only the forward chaining that is found by such backward chaining that changes probabilities. Am I correct in this interpretation of what Abram said, and is that interpretation included in what your Google clippings indicate is the generally understood meaning of the term backward chaining. Ed Porter P.S. I would appreciate answers for Abram or any else on this list who understands the question and has some knowledge on the subject. -Original Message- From: Lukasz Stafiniak [mailto:[EMAIL PROTECTED] Sent: Tuesday, July 15, 2008 3:05 AM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? On Tue, Jul 15, 2008 at 8:01 AM, Brad Paulsen [EMAIL PROTECTED] wrote: The terms forward-chaining and backward-chaining when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate if/then rules in a rule base (RB). In the FWC algorithm, the if part is evaluated and, if TRUE, the then part is added to the FWC engine's output. In the BWC algorithm, the then part is evaluated and, if TRUE, the if part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. Curiously, the definition put by Abram Demski is the only one I've been aware of until yesterday (I believe it's the one used among theorem proving people). Let's see what googling says on forward chaining: 1. (Wikipedia) 2. http://www.amzi.com/ExpertSystemsInProlog/05forward.htm A large number of expert systems require the use of forward chaining, or data driven inference. [...] Data driven expert systems are different from the goal driven, or backward chaining systems seen in the previous chapters. The goal driven approach is practical when there are a reasonable number of possible final answers, as in the case of a diagnostic or identification system. The system methodically tries to prove or disprove each possible answer, gathering the needed information as it goes. The data driven approach is practical when combinatorial explosion creates a seemingly infinite number of possible right answers, such as possible configurations of a machine. 3. http://ai.eecs.umich.edu/cogarch0/common/prop/chain.html Forward-chaining implies that upon assertion of new knowledge, all relevant inductive and deductive rules are fired exhaustively, effectively making all knowledge about the current state explicit within the state. Forward chaining may be regarded as progress from a known state (the original knowledge) towards a goal state(s). Backward-chaining by an architecture means that no rules are fired upon assertion of new knowledge. When an unknown predicate about a known piece of knowledge is detected in an operator's condition list, all rules relevant to the knowledge in question are fired until the question is answered or until quiescence. Thus, backward chaining systems normally work from a goal state back to the original state. 4. http://www.ontotext.com/inference/reasoning_strategies.html * Forward-chaining: to start from the known facts and to perform the inference in an inductive fashion. This kind of reasoning can have diverse objectives, for instance: to compute the inferred closure; to answer a particular query; to infer a particular sort of knowledge (e.g. the class taxonomy); etc. * Backward-chaining: to start from a particular fact or from a query and by means of using deductive reasoning to try to verify that fact or to obtain all possible results of the query. Typically, the reasoner decomposes the fact into simpler facts that can be found in the knowledge base or transforms it into alternative facts that can be proven applying further
RE: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Jim, Sorry. Obviously I did not understand you. Ed Porter -Original Message- From: Jim Bromer [mailto:[EMAIL PROTECTED] Sent: Tuesday, July 15, 2008 9:33 AM To: agi@v2.listbox.com Subject: RE: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? Ed Porter said: You imply you have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimensional and presumably large semantic space. Unfortunately I was unable to understand from your description how you claimed to have accomplished this. - I never implied that I have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimension and presumably large semantic space. I clearly stated: I have often talked about the use of multi-level complex methods and I see some similarity to the ideas that they discussed to my ideas. -and, The complex groupings of objects that I have in mind would have been derived using different methods of analysis and combination and when a group of them is called from an input analysis their use should tend to narrow the objects that might be expected given the detection by the feature detectors. Although I haven't expressed myself very clearly, this is very similar to what Riesenhuber and Poggio were suggesting that their methods would be capable of. So, yes,I think some similar methods can be used in NLP. I clearly used the expression in mind just to avoid the kind of misunderstanding that you made. I never made the exaggerated claim that I had accomplished it. The difference between having an idea in mind and having claimed to have accomplished a goal, which the majority of participants in the group would acknowledge is elusive, should be obvious and easy to understand. I am not claiming that I have a method that would work in all semantic space. I would be happy to claim that I do have a theory which I believe should show some limited extensibility in semantic space that goes beyond other current theories. However, I will not know for sure until I test it and right now that looks like it would be years off. I would be happy to continue the dialog if it can be conducted in a less confrontational and more genial manner than it has been during the past week. Jim Bromer Jim, In the Riesenhuber and Poggio paper the binding that were handled implicitly involved spatial relationships, such as an observed roughly horizontal line substantially touching an observed roughly vertical line at their respective ends, even though their might be other horizontal and vertical lines not having this relationship in the input pixel space. It achieves such implicit bindings by having enough separate models to be able to detect, by direct mapping, such a touching relationship between a horizontal and vertical lines at each of many different locations in the visual input space. But the Poggio paper deals with a relatively small number of relationships in a relatively small (160x160) low dimensional (2d) space using 23 million models. You imply you have been able to accomplish a somewhat similar implicit representation of bindings in a much higher dimensional and presumably large semantic space. Unfortunately I was unable to understand from your description how you claimed to have accomplished this. Could you please clarify you description with regard to this point. Ed Porter -Original Message- From: Jim Bromer [mailto:[EMAIL PROTECTED] Sent: Monday, July 14, 2008 1:38 PM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? I started reading a Riesenhuber and Poggio paper and there are some similarities to ideas that I have considered although my ideas were explicitly developed about computer programs that would use symbolic information and are not neural theories. It is interesting that Risesnhuber and Poggio argued that the binding problem seems to be a problem for only some models of object recognition. In other words, it seems that they are claiming that the problem disappears with their model of neural cognition! The study of feature detectors in cats eyes is old news and I did incorporate that information into the development of my own theories. I have often talked about the use of multi-level complex methods and I see some similarity to the ideas that they discussed to my ideas. In my model an input would be scanned for different features using different kinds of analysis on the input. So then a configuration of simple features would be derived from the scan and these could be associated with a number of complex groups of objects that have been previously associated with the features. Because the complex groups of objects are complexes (in the general sense), and would be learned by previous experience, they are not insipidly modeled on one
Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Am I correct in this interpretation of what Abram said, and is that interpretation included in what your Google clippings indicate is the generally understood meaning of the term backward chaining. Ed Porter It sounds to me like you are interpreting me correctly. One important note. Lukasz quoted one source that claimed that forward chaining can help to cut down the combinatorial explosion arising from the huge search space in backwards-chaining. This is true in some situations, but the opposite can also be the case; backwards-chaining can help to focus inferences when it would be impossible to deduce every fact that would follow by forward-chaining. It depends on the forward and backwards branching factors. If every fact fires an average of five rules forwards, but three backwards, then backwards-chaining will be less expensive; 5^n vs 3^n, where n is the length of the actual deductive chain being searched for. Simultaneous backwards/forwards chaining that meets in the middle can be even less expensive; with a branching factor of 2 in both directions, the search time goes down from 2^n for forward or backward chaining to 2^(n/2 + 1). On the other hand, what we want the system to do makes a big difference. If we really do have a single goal-sentence we want to prove or disprove, the above arguments hold. But if we want to deduce all consequences of our current knowledge, we should use forward chaining regardless of branching factors and so on. Most of this stuff should be in any intro AI textbook. --Abram On Tue, Jul 15, 2008 at 11:08 AM, Ed Porter [EMAIL PROTECTED] wrote: Lukasz, Your post below was great. Your clippings from Google confirm much of the understanding that Abram Demski was helping me reach yesterday. In one of his posts Abram was discussing my prior statement that top-down activation could be either forward or backward chaining. He said If the network is passing down an expectation based on other data, informing the lower network of what to expect, then this is forward chaining. But if the signal is not an expectation, but more like a query pay attention to data that might conform/contradict this hypothesis, and notify me ASAP then it is backwards chaining. And it seems realistic that it can be both of these. I am interpreting this quoted statement as implying the purpose of backward chaining is to search for forward chaining paths that either confirm or contradict a pattern of interest or that provide a path or plan to a desired goal. In this view the backward part of backward chaining provides no changes in probability, only changes in attention, and it is only the forward chaining that is found by such backward chaining that changes probabilities. Am I correct in this interpretation of what Abram said, and is that interpretation included in what your Google clippings indicate is the generally understood meaning of the term backward chaining. Ed Porter P.S. I would appreciate answers for Abram or any else on this list who understands the question and has some knowledge on the subject. -Original Message- From: Lukasz Stafiniak [mailto:[EMAIL PROTECTED] Sent: Tuesday, July 15, 2008 3:05 AM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? On Tue, Jul 15, 2008 at 8:01 AM, Brad Paulsen [EMAIL PROTECTED] wrote: The terms forward-chaining and backward-chaining when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate if/then rules in a rule base (RB). In the FWC algorithm, the if part is evaluated and, if TRUE, the then part is added to the FWC engine's output. In the BWC algorithm, the then part is evaluated and, if TRUE, the if part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. Curiously, the definition put by Abram Demski is the only one I've been aware of until yesterday (I believe it's the one used among theorem proving people). Let's see what googling says on forward chaining: 1. (Wikipedia) 2. http://www.amzi.com/ExpertSystemsInProlog/05forward.htm A large number of expert systems require the use of forward chaining, or data driven inference. [...] Data driven expert systems are different from the goal driven, or backward chaining systems seen in the previous chapters. The goal driven approach is practical when there are a reasonable number of possible final answers, as in the case of a diagnostic or identification system. The system methodically tries to prove or disprove each possible answer, gathering the needed information as it goes. The data driven approach is practical when combinatorial explosion creates a seemingly infinite number
RE: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
Abram, Thanks, for the info. The concept that the only purpose of backward chaining to find appropriate forward chaining paths, is an important clarification of my understanding. Ed Porter -Original Message- From: Abram Demski [mailto:[EMAIL PROTECTED] Sent: Tuesday, July 15, 2008 11:38 AM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? Am I correct in this interpretation of what Abram said, and is that interpretation included in what your Google clippings indicate is the generally understood meaning of the term backward chaining. Ed Porter It sounds to me like you are interpreting me correctly. One important note. Lukasz quoted one source that claimed that forward chaining can help to cut down the combinatorial explosion arising from the huge search space in backwards-chaining. This is true in some situations, but the opposite can also be the case; backwards-chaining can help to focus inferences when it would be impossible to deduce every fact that would follow by forward-chaining. It depends on the forward and backwards branching factors. If every fact fires an average of five rules forwards, but three backwards, then backwards-chaining will be less expensive; 5^n vs 3^n, where n is the length of the actual deductive chain being searched for. Simultaneous backwards/forwards chaining that meets in the middle can be even less expensive; with a branching factor of 2 in both directions, the search time goes down from 2^n for forward or backward chaining to 2^(n/2 + 1). On the other hand, what we want the system to do makes a big difference. If we really do have a single goal-sentence we want to prove or disprove, the above arguments hold. But if we want to deduce all consequences of our current knowledge, we should use forward chaining regardless of branching factors and so on. Most of this stuff should be in any intro AI textbook. --Abram On Tue, Jul 15, 2008 at 11:08 AM, Ed Porter [EMAIL PROTECTED] wrote: Lukasz, Your post below was great. Your clippings from Google confirm much of the understanding that Abram Demski was helping me reach yesterday. In one of his posts Abram was discussing my prior statement that top-down activation could be either forward or backward chaining. He said If the network is passing down an expectation based on other data, informing the lower network of what to expect, then this is forward chaining. But if the signal is not an expectation, but more like a query pay attention to data that might conform/contradict this hypothesis, and notify me ASAP then it is backwards chaining. And it seems realistic that it can be both of these. I am interpreting this quoted statement as implying the purpose of backward chaining is to search for forward chaining paths that either confirm or contradict a pattern of interest or that provide a path or plan to a desired goal. In this view the backward part of backward chaining provides no changes in probability, only changes in attention, and it is only the forward chaining that is found by such backward chaining that changes probabilities. Am I correct in this interpretation of what Abram said, and is that interpretation included in what your Google clippings indicate is the generally understood meaning of the term backward chaining. Ed Porter P.S. I would appreciate answers for Abram or any else on this list who understands the question and has some knowledge on the subject. -Original Message- From: Lukasz Stafiniak [mailto:[EMAIL PROTECTED] Sent: Tuesday, July 15, 2008 3:05 AM To: agi@v2.listbox.com Subject: Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? On Tue, Jul 15, 2008 at 8:01 AM, Brad Paulsen [EMAIL PROTECTED] wrote: The terms forward-chaining and backward-chaining when used to refer to reasoning strategies have absolutely nothing to do with temporal dependencies or levels of reasoning. These two terms refer simply, and only, to the algorithms used to evaluate if/then rules in a rule base (RB). In the FWC algorithm, the if part is evaluated and, if TRUE, the then part is added to the FWC engine's output. In the BWC algorithm, the then part is evaluated and, if TRUE, the if part is added to the BWC engine's output. It is rare, but some systems use both FWC and BWC. That's it. Period. No other denotations or connotations apply. Curiously, the definition put by Abram Demski is the only one I've been aware of until yesterday (I believe it's the one used among theorem proving people). Let's see what googling says on forward chaining: 1. (Wikipedia) 2. http://www.amzi.com/ExpertSystemsInProlog/05forward.htm A large number of expert systems require the use of forward chaining, or data driven inference. [...] Data driven expert systems are different from the goal driven, or backward chaining systems
Re: FW: [agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
4. http://www.ontotext.com/inference/reasoning_strategies.html * Forward-chaining: to start from the known facts and to perform the inference in an inductive fashion. This kind of reasoning can have diverse objectives, for instance: to compute the inferred closure; to answer a particular query; to infer a particular sort of knowledge (e.g. the class taxonomy); etc. * Backward-chaining: to start from a particular fact or from a query and by means of using deductive reasoning to try to verify that fact or to obtain all possible results of the query. Typically, the reasoner decomposes the fact into simpler facts that can be found in the knowledge base or transforms it into alternative facts that can be proven applying further recursive transformations. A system like Clips is forward chaining but there is no induction going on. Whether fwd- or bkw- chaining it is deduction as far as I've ever heard of. With induction we are implying repeated observations that lead to some new knowledge (ie., some new rule in this case). That was my understanding anyway, and I'm no PhD scientist. Mike Archbold --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
