jps@cat.cmu.edu (James Salsman) (08/07/89)
In article <9795@phoenix.Princeton.EDU> harnad@phoenix.Princeton.EDU (Stevan Harnad) writes: > James Salsman (jps@cat.cmu.edu) wrote: >> THE SYMBOL GROUNDING PROBLEM... No problem: >> >> The | /"Iconic"\ Distributed / \ / \Predicates &/ >> Out |--------< Buffer >-------------< EPAM >------< SOAR >----------<LTM >> side| Signals \ STM / Representation\ / Symbols\ / Productions\ > This is exactly the kind of naive "hook-em-all-up-together" modularism that I > wrote the preprint in question in order to refute. Now one can either hold > onto these unexamined beliefs (and the dead-ends they lead to in cognitive > modeling) or wake up and smell the coffee... Some prefer iced tea. The theories shown in my diagram could hardly be considered unexamined. The EPAM and SOAR models have both withstood the test of several years of scientific scrutiny both in the classrooms of our nation's Universities and in the peer review process of international journals. EPAM (Elementary Perciever And Memorizer) was constructed by Herbert Simon and Edward Figenbaum in the late 1950's. Formally, EPAM is an algorithm that maps sensory signals into symbolic information. In addition, the EPAM model has been shown to predict the behavior of humans in several cognitive experimental domains such as response time measurements, patterns of forgetting, and relation of the serial position of sensory information to those factors. Additionally, EPAM, as a theory, has been described in terms of an algorithm and a data structure, so it can be implemented in practice as a program for either a parallel or serial computer. EPAM is best described as a "discrimination network" with some very special features. [For a good introduction see "The Handbook of Artificial Intelligence," Cohen, P., Figenbaum, E., 1982, William Kaufman, Inc., Volume III, Section XI.D.] A Production System is a rule-based programming language that seems to be a good model of the lower-brain conditioning systems (S --> R). Allen Newell has constructed a descendant of GPS called SOAR that explains higher-level cognition in terms of a Production System (in practice, SOAR is written on top of Forgy's OPS5.) SOAR is a goal-directed general problem solving system with an experience-based learning model (Chunking), a set of powerful default search routines implemented as "weak methods," and a way to generate useful subgoals when the system reaches an impasse. Production Systems are "forward chaining" inference systems, as opposed to the "backward chaining" systems of Prolog -- this does not mean that production systems are incapable of logical inference -- only that they have a choice of several methods with varing degrees of effectiveness for different problems, including Prolog's Depth-First-Search. The literature correlating SOAR's behavior with that of humans is based on extensive protocol and timing analysis and is especially obvious in areas such as stimulus-response compatibility and the shapes of the leaning curves of several different kinds of cognitive problems. It is clear that Connectionist Neural-Network and Parallel Distributed Processing research, present a very rich description of the "computing machinery" of the brain. We have reached the point in our understanding of cognition that we may now begin the implentation of programs such as "the mind" on computing machinery other than the brain. I've already put together a workable dataflow PS on the Connection Machine, but the SIMD architecture didn't allow enough state per processing element to properly handle the binding semantics of OPS5, which SOAR is heavly dependant on. I am currently working on a distributed production system using the Transputer's MIMD multiprocessor architecture that will include all of OPS5's binding semantics. On a 16-Transputer system with a hypercube topology and 256K RAM per processor, I expect to be able to construct a system with a capacity of 40000 total productions and still maintain a 10 millisecond cycle time. The bad news is that this will require Transputer machine code generation while the system is running -- hardly pure OCCAM. The 10 ms figure has been specified by both Newell and Simon as the assumed firing time for their human cognition models. So, to return to the symbol grounding problem: SOAR is a purely symbolist architecture, and requires another underlying system to parse sensory signals into symbolic form. That is exactly the function of EPAM. The combined system should be able perform real-time cognitive tasks such as vocal communication, perhaps as a natural language processing back-end for the Sphinx real-time speaker-independent speech recogntion system that CMU has constructed. Such a system would make a very freindly user interface for an information retrival network (like the Telephone system.) At Carnegie-Mellon, we do not use "hook-em-all-up-together" strategies for cognitive modeling. We use "functional composition," instead. Aren't there any hackers at Princeton? Or are you all "Software Engineers?" Either way, it should be easy for you to understand the difference. :James Disclaimer: I have worked for the SOAR project, but I don't represent any official viewpoint. -- :James P. Salsman (jps@CAT.CMU.EDU)
harnad@elbereth.rutgers.edu (Stevan Harnad) (08/08/89)
jps@cat.cmu.edu (James Salsman) of Carnegie Mellon wrote: > The theories shown in my diagram could hardly be considered unexamined... > Formally, EPAM is an algorithm that maps sensory signals into symbolic > information... [It] predict[s] the behavior of humans in several > cognitive experimental domains such as response time measurements, > patterns of forgetting, and relation of the serial position of sensory > information to those factors... literature correlating SOAR's > behavior with that of humans is based on extensive protocol and timing > analysis and is especially obvious in areas such as stimulus-response > compatibility and the shapes of the learning curves of several > different kinds of cognitive problems. It's not the theoretical modules themselves that are unexamined, but the belief that hooking them together will successfully model the mind. Fitting the fine-tuning parameters of toy-size pieces of performance is not the mark or measure of a successful model of the mind. Life-size performance capacity (never mind its fine-tuning for now) is. The test of whether an algorithm "maps sensory signals into symbolic information" in the right way is to see whether it produces human-scale robotic performance, including whether your "cat is on the mat" symbol string, along with all the other life-size things it does, really picks out cats, mats, and the one being on the other. The symbol grounding problem is a harbinger of the possibility that neither the purely symbolic approach alone, nor a simple hook-up between pure symbol systems and transducers or other autonomous functional modules, can lead to a successful lifesize performance model. > Connectionist Neural-Network and Parallel Distributed Processing > research presents a very rich description of the "computing machinery" > of the brain. We have reached the point in our understanding of > cognition that we may now begin the implentation of programs such as > "the mind" on computing machinery other than the brain. Unfortunately, the [unexamined] point at issue in the symbol grounding problem is the very notion that the mind is a program (i.e., symbol manipulation) and that the brain is just the machine on which it's implemented. (There are, by the way, more perspicuous potential uses for connectionism in modeling the mind than simply as one more way to implement symbol-manipulation.) > binding semantics of OPS5... Grounding semantics and binding semantics are unfortunately not the same thing. > I expect to be able to construct a system with a capacity of 40000 > total productions and still maintain a 10 millisecond cycle time... > [which] has been specified by both Newell and Simon as the assumed > firing time for their human cognition models. [See discussion of fine-tuning parameters above. Same goes for numerology. It might be better to focus your expectations instead in the direction of generating lifesize performance capacity; I'm afraid that that's so much more than 40000 productions that it makes it unlikely that the way to scale up is simply to add more productions...] > SOAR is a purely symbolist architecture, and requires another > underlying system to parse sensory signals into symbolic form. That is > exactly the function of EPAM. The combined system should be able to > perform real-time cognitive tasks such as vocal communication, perhaps > as a natural language processing back-end for the Sphinx real-time > speaker-independent speech recogntion system that CMU has constructed. > Such a system would make a very friendly user interface for an > information retrieval network (like the Telephone system.)... > At Carnegie-Mellon, we do not use "hook-em-all-up-together" strategies > for cognitive modeling. We use "functional composition," instead. These "back-ends" and "interfaces" sure sound like hook-ups to me; but call it functional composition if you like. (Maybe then the dedicated hybrid nonsymbolic/symbolic grounding system I'm proposing is just a matter of "functional composition" too...) What you have to decide, however, is whether you're doing information retrieval aids for people (in which case you don't have to worry about grounding your symbols, because the [grounded] users can interpret them for themselves) or you're modeling the mind -- in which case the meanings of the symbols will have to be intrinsic, not parasitic on human interpreters. > Aren't there any hackers at Princeton? Or are you all "Software Engineers?" First, my views do not represent those at Princeton (or Rutgers, as the case may be). Second, it is only from a certain vantage point that psychological theory is seen as "software engineering"... -- Stevan Harnad INTERNET: harnad@confidence.princeton.edu harnad@princeton.edu srh@flash.bellcore.com harnad@elbereth.rutgers.edu harnad@princeton.uucp BITNET: harnad@pucc.bitnet CSNET: harnad%princeton.edu@relay.cs.net (609)-921-7771