neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") (08/29/89)
Neuron Digest Monday, 28 Aug 1989 Volume 5 : Issue 36 Today's Topics: CG Methods Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Re: Connectionism, a paradigm shift? Paradigm Shift Response (sort of) Re: Paradigm Shift Response (sort of) Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" Use "ftp" to get old issues from hplpm.hpl.hp.com (15.255.176.205). ------------------------------------------------------------ Subject: CG Methods From: tedwards@cmsun.nrl.navy.mil (Thomas Edwards) Date: Mon, 28 Aug 89 13:08:38 -0400 (Many people have written to me for more information on this conjuagte gradient reference. Since I know mail has bounced from me to alot of people who wrote, here it is:) The reference is: Kramer, A. and Sangiovanni-Vincentelli, A. "Efficient Parallel Learning Algorithms for Neural Networks." _Advances in Neural Information Processing Systems I_, ed. D. Touretzky. Morgan Kaufmann Publishers, Inc. San Mateo, Ca. 1989. ISBN 1-558-60015-9 This article discusses backprop, steepest descent, and conjugate-gradient (Polak-Ribiere rule, actually a good discussion of the rule, but no serious discussion of the line minimization except a reference to Luenberger) methods on The Connection Machine. Results are reported, but actual parallel data representation and computation not well discussed. I have been examining steepest descent and conjugate gradient methods. I am finding that there are often times when the point of line minimization is so close to the current weight point in the search direction that the line search has serious trouble not finding a "minimum" which is actually much worse than the current position. The line search is no doubt the most difficult part of conjugate gradient and steepest descent programs. Finding the best original points for the line search to use can be a chore. I have yet to perform a steepest descent or conjugate gradient search which is computationally more efficient than backpropagation, but I know from using other conjugate gradient programs that it is possible, although alot of work (and mathematical theory) has to go into the program design. Perhaps that is why conjugate gradient learning has not been explored as much as backpropagation. -Thomas Edwards tedwards@cmsun.nrl.navy.mil ins_atge@jhunix.hcf.jhu.edu ins_atge@jhuvms.BITNET ------------------------------ Subject: Connectionism, a paradigm shift? From: dave@cogsci.indiana.edu (David Chalmers) Organization: Indiana University, Bloomington Date: Thu, 03 Aug 89 23:40:44 +0000 [[ Editor's Note: The following [edited] discussion started on the AI bulletin board and appeared on the USENET group. While much of the discussion in the Digest is technical, we often need to take stock in the broader question of what we're doing and why. The parallels between the AI craze and the current Neural Network phenonmena are too easy to make. Shoud we believe our own PR? -PM ]] Almost half the papers at this month's upcoming Cognitive Science conference are about connectionism! For a field which just 3-4 years ago was very small and the "radical new kid on the block," this is an amazing growth. Of course, there has been much talk of a "paradigm shift." But paradigm shifts were never meant to happen this fast. The electronic age seems to accelerate everything (remember cold fusion?). There's no chance for a slow, graceful growth in the field; the bandwagon has arrived and it's moving fast, jump on before it's too late! This unnatural acceleration has got to lead to unstable, unpredictable consequences. There seems to be already almost as much valueless work in connectionism as there was in "traditional AI" (tweak this, try that, apply here, generalize there, and quickly, before somebody else does!). Prediction: within a year or two an "anti-connectionist" backlash will be growing very prominent. (There are already a few signs.) After all the hype, people will begin to grumble "come on, they're just smart pattern recognizers/associators. Can they really do _cognition_?" In this accelerated age, these views will quickly become conventional wisdom, and many will jump off the bandwagon as quickly as they jumped on. Meantime, in the background people will keep plugging away, doing good connectionist work at the slow and steady pace that good science seems to require. My prediction: in the wash, connectionism (along with other "emergent" approaches) will emerge as the dominant and most successful paradigm, but not for another decade yet, and not before another couple of violent swings in various directions. Comments? Dave Chalmers (dave@cogsci.indiana.edu) Concepts and Cognition, Indiana University. "Whereof one cannot speak, thereof one must make it all up." ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: andrew@berlioz (Lord Snooty @ The Giant Poisoned Electric Head ) Organization: National Semiconductor, Santa Clara Date: Fri, 04 Aug 89 02:00:38 +0000 In article <24241@iuvax.cs.indiana.edu>, dave@cogsci.indiana.edu (David Chalmers) writes: [A discussion on fads and the rapid growth of connectionism, and a prediction of its demise through hype] I think you should crosspost this to comp.ai.neural-nets, whose members seem to exhibit the usual healthy cynicism of a comp.. group; not a pack of zealots by any means! I agree that there exists a danger from over-rapid over-exposure and the concomitant media hype. This is a constant warning cry made at the conferences, and by people who popularise the field. You have to bear in mind that we're only human, and become naturally excited even as researchers and informed observers when new results appear. It is not necessary to *immediately* understand the nature of the underlying mechanism when a new and successful application is created (in this sense, your analogy to cold fusion is spot-on). I think that what is required to save the field from the "hype seesaw" is a healthy rate of generation of solid new theoretical results. Two fairly recent results, for example, which could be seen to qualify: 1) A preprocessing paradigm using a simple one-layer net and an easily- implementable learning algorithm, which extracts the eigenvalues of the input autocorrelation - useful for image compression, etc. In particular, information-theoretic approaches are producing new results. [Sanger, Linsker, Foldiak] 2) A formal proof of an algorithm for a restricted class of nets, which predicts detailed network dynamics given the training pattern set. [Lemmon, Kumar] There is a tremendous amount of high-quality work going on, bolstered by the application of formal mathematical techniques. It seems to me that this truly sets NN research apart from the much more "hand-waving" stuff that I encountered when looking at conventional AI, when expert systems were on the rise in the early- and mid-80s. Here one found tree traversal stuff and Bayesian statistical variations, definitons of "frames" and the like; the ad hoc component was significant. (although fuzzy set theory has to some extent set some of this on a more formal footing, I have to agree). The analogy I have in mind equates NN research to the microstructure of cognition, and as such is akin to "physics". When dealing with the atoms of behaviour, it's possible to produce significant and fundamental results. Symbolic AI smacks to me much more like "inorganic chemistry". The consensus view seems to be that these two paradigms will eventually cooperate in future artificial cognitive systems. Work is already ongoing to combine expert systems with NN coprocessors. However, taking the brain as an existence proof, it's clear that NN technology can implement all levels of cognition, whereas it is unclear whether symbolic methods are capable of this [see e.g. Steve Harnad: subsymbolic and symbolic processing]. ........................................................................... Andrew Palfreyman There's a good time coming, be it ever so far away, andrew@berlioz.nsc.com That's what I says to myself, says I, time sucks jolly good luck, hooray! ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: dmark@cs.Buffalo.EDU (David Mark) Organization: SUNY/Buffalo Geography Date: Sat, 05 Aug 89 12:54:24 +0000 I agree that it does look like a paradigm shift, since it is a radically new way to look at some problems. I have not yet become very interested in connectionist models, because, as a scientist rather than an engineer, I am interested primarily in seeking _explanation_ rather than _performance_. There is little doubt that NN programs based on the connectionist paradigm perform some computing tasks very well, including some (many) tasks of an AI/ES flavor. But, I am not aware of a lot of success in understanding what the weights _MEAN_, except for some specialized fields such as low-level vision work, in which we also have neurophysiological evidence. Now, I only read a small proportion of the NN/connectionist work, so I wonder if _explanations_ using NN/C have become more eveident in the last year or two. If not, I'm not interested, and assume that the "old" paradigm will remain quite healthy in the sciences at least. (Just my underinformed opinions, obviously not necessarily those of my colleagues!) David Mark dmark@cs.buffalo.edu ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: holt@cs.AthabascaU.CA (Peter Holt) Organization: Athabasca U, Alberta, Canada Date: Mon, 07 Aug 89 14:46:04 +0000 I would say that is a fairly zealous statement. Personally I have not decided which paradigm is better for what when yet, but lets remember that there may only be a superficial resemblance between the operations of the brain and current neural net technology! A lot more things are happening in the brain (especially chemically and at the intraneuron level) than are in neural nets. It may even be a coincidence that what some of the functionality of neural nets approximates some of the very basic perceptual-cognitive functions of the brain. Some of the other functionality of neural nets (extracting eigenvalues?) would not seem to match the way humans do same things at all. ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: andrew@berlioz (Lord Snooty @ The Giant Poisoned Electric Head ) Organization: National Semiconductor, Santa Clara Date: Mon, 07 Aug 89 19:59:55 +0000 In article <705@aurora.AthabascaU.CA>,holt@cs.AthabascaU.CA (Peter Holt) writes: > ..but lets remember that there may only be a superficial resemblance > between the operations of the brain and current neural net technology! Let's talk about "resemblance", then. "Resemblance" is a strong suit for nets in the connectionism vs. serial symbolic systems debate, and yet you use it for critique! When PROLOG executes a branch instruction in the ALU of the SPARC chip, where is the resemblance to the brain? ........................................................................... Andrew Palfreyman There's a good time coming, be it ever so far away, andrew@berlioz.nsc.com That's what I says to myself, says I, time sucks jolly good luck, hooray! ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: jps@cat.cmu.edu (James Salsman) Organization: Carnegie Mellon Date: Tue, 08 Aug 89 03:15:17 +0000 > When PROLOG executes a branch instruction in the ALU of the SPARC chip, > where is the resemblance to the brain? It depends on the rest of the SPARC system's state. If you have a formal description of a data structure and an algorithm, then you have a program. Using a technique called "programming" one may map these descriptions on to different kinds of computer systems. The Neural-Net of the brain is one kind of system, and a SPARC system is somthing else entirely. The only reason that they can't be executing the same program is that the I/O systems are very different. :James P. Salsman (jps@CAT.CMU.EDU) ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: coggins@coggins.cs.unc.edu (Dr. James Coggins) Organization: University Of North Carolina, Chapel Hill Date: Sun, 13 Aug 89 14:34:01 +0000 [[ Editor's Note: the beginning of this message has been cropped. -PM ]] >There is a tremendous amount of high-quality work going on, bolstered by >the application of formal mathematical techniques. I'm afraid that the theoretical foundation you appreciate is actually inherited (or bastardized, depending on your point of view) from the statistical pattern recognition studies of ten to twenty years ago. Sure there is a theory base, but it's ready-made, much of it not arising inherently from NNs (but being REdiscovered there). "...only be sure please always to call it RESEARCH!" from Lobachevsky by Tom Lehrer I have been impressed with the confirmation provided by this newsgroup that the majority of researchers in this area really are disgusted at the publicity-mongering, money-grubbing approach of too many well-placed (and well-heeled) labs, researchers, writers, companies, seminar sellers, and the like. NNs might become a significant contribution making possible highly parallel implementations of many kinds of processes if the science fiction futurist brain-theory dabblers would shut up and let the real researchers develop the field in a careful, disciplined way, without having to run interference against massively inflated expectations of the work. A few months ago I posted to comp.ai.neural-nets the document reproduced below. I guess it was too hot for the newsgroup, but I did receive 13 e-mail replies: 8 firmly supportive, 4 asking for more pointers to statistical pattern recognition which I gladly supplied (But note: Is the scholarship in the NN field really so weak that NN researchers are unaware of twenty years of research in statistical pattern recognition? The evidence says yes!), and one sharply critical but easy to refute (a True Believer who went down in flames). I posted the document below in the spirit of my other "Outrageous Discussion Papers" that I have been circulating to carefully selected audiences to provoke thought and comment and encourage skepticism. I have one flaming the use of rule-based expert systems in medical applications, one arguing that edges are an inadequate foundation for vision, one arguing that automatic identification of organs in CT scans is an unworthy task of little practical value, one that is a manifesto for my approach to computer vision research, and the neural net one below. If you are interested, e-mail me, but I'm leaving now for a three-week vacation, so don't expect my usual rapid response. --------------------------------------------- My assessment of the neural net area is as follows: (consider these Six Theses nailed to the church door) 1. NNs are a parallel implementation technique that shows promise for making perceptual processes run in real time. 2. There is nothing in the NN work that is fundamentally new except as a fast implementation. Their ability to learn incrementally from a series of samples nice but not new. The way they learn and make decisions is decades old and first arose in communication theory, then was further developed in statistical pattern recognition. 3. The claims that NNs are fundamentally new are founded on ignorance of statistical pattern recognition or on simplistic views of the nature of statistical pattern recognition. I have heard supposedly competent people working in NNs claim that statistical pattern recognition is based on assumptions of Gaussian distributions which are not required in NNs, therefore NNs are fundamentally different. This is ridiculous. Statistical pattern recognition is not bound to Gaussians, and NNs do, most assuredly, incorporate distributional assumptions in their decision criteria. 4. A more cynical view that I do not fully embrace says that the main function of "Neural Networks" is as a label for money. It is a flag you wave to attract money dispensed by people who are interested in the engineering of real-time perceptual processing and who are ignorant of statistical pattern recognition and therefore the lack of substance of the neural net field. 5. Neural nets raise lots of engineering questions but little science. Much of the excitement they have raised is based on uncritical acceptance of "neat" demos and ignorance. As such, the area resembles a religion more than a science. 6. The "popularity" of neural net research is a consequence of the miserable mathematical backgrounds of computer science students (and some professors!). You don't need to know any math to be a hacker, but you have to know math and statistics to work in statistical pattern recognition. Thus, generations of computer science students are susceptible to hoodwinking by neat demos based on simple mathematical and statistical techniques that incorporate some engineering hacks that can be tweaked forever. They'll think they are accomplishing something by their endless tweaking because they don't know enough math and statistics to tell what's really going on. Dr. James M. Coggins coggins@cs.unc.edu Computer Science Department A neuromorphic minimum distance classifier! UNC-Chapel Hill Big freaking hairy deal. Chapel Hill, NC 27599-3175 -Garfield the Cat and NASA Center of Excellence in Space Data and Information Science ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: lee@uhccux.uhcc.hawaii.edu (Greg Lee) Organization: University of Hawaii Date: Sun, 13 Aug 89 20:18:55 +0000 >5. Neural nets raise lots of engineering questions but little science. Judging from popular accounts, and as an outsider to the field, this is the impression I get -- that NNs are an attempt to do technology without science. I have seen what I take to be kindred approaches in my own field, linguistics. The idea seems to be that one can escape the necessity to achieve an understanding of human perception and leave that to a machine (or algorithm, rather). Since scientific understanding (new and old) is so difficult to come by, it's a very seductive idea. But not a reasonable one. Greg, lee@uhccux.uhcc.hawaii.edu ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: bph@buengc.BU.EDU (Blair P. Houghton) Organization: Boston Univ. Col. of Eng. Date: Mon, 14 Aug 89 01:02:33 +0000 > >>5. Neural nets raise lots of engineering questions but little science. Eh? Science has been the forming of models and the fitting of them to observed phenomena. In the case of artificial neural systems, the models are physical entities (neuromimes, simulations of neuromimes, simulations of behavioral models of neuromimes and of elements composed of neuromimes, etc.) rather than tautologies (laws, theorems, etc.), and the fit is a behavioral one, as is every theory, until a new, deeper observation is made of the behavior, or until we are prepared to discard degenerative assumptions that limit our study of currently observed behavior. >The idea seems to be that one can escape the >necessity to achieve an understanding of human perception and leave that >to a machine (or algorithm, rather). Since scientific understanding >(new and old) is so difficult to come by, it's a very seductive idea. >But not a reasonable one. I seem to remember having this same conversation before...anyway: Doing neural nets this way is akin to allowing probability to be a mathematical field, and to statistical mechanics and quantum theory. The understanding has, and consciously so, been behind the techniques in those areas since the techniques were first found to be superior to the understanding in predictive power. --Blair "It's quite reasonable. It's quite reasonable to assume that my thesis won't be half this erudite." ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: ari@kolmogorov.physics.uiuc.edu Date: Tue, 15 Aug 89 01:16:00 +0000 Much of the hype with Neural Networks sounds much like the hype in the study of Chaos. One author of a popular book on Chaos claims a paradigm shift in physical science, even as far as to claim that the 20th century will be remembered for the theory of General Relativity, Quantum Mechanics and the theory of Chaos! One difficulty in the field of Chaos is the mixing of hype with solid theoretical and conceptual advances. Chaos is a broad title given to a large class of ideas and observed (usually computationally) phenomenon as well as some theory. It is much more a collection of bits and pieces and tantalizing glimpses than a cohesive theory. One posting claims that: "Doing neural nets this way is akin to allowing probability to be a mathematical field, and to statistical mechanics and quantum theory." Which seems to imply that the fields of probability, statistical mechanics (my own field) and quantum theory are in some sense the less precise version of some other field or fields which simply simulate, rather than theorize. These views seem wrong to me, and certainly, the bulk of NN research appears to be at a much less about theory, and much more about description and simulation. This is very well and good, and is much more akin to Monte Carlo Ising Spin simulations in statistical physics. However, such simulations are not the bulk of statistical physics. The current legacy of Chaos theory is a more descriptive rather than theoretical understanding of chaotic phenomena. Of course solid work has been done, but a lot of pretty pictures have made more than the fair share of impact. I believe it is important to any field to understand the differences between observing, describing, classifying and understanding phenomena. One should not claim the last simply from the first. Aritomo Shinozaki co/ Physical Theory Group ari@kolmogorov.physics.uiuc.edu Beckman Institute University of Illinois, Urbana-Champaign Urbana IL, 61801 ------------------------------ Subject: Re: Connectionism, a paradigm shift? From: jtn@potomac.ads.com (John T. Nelson) Organization: Advanced Decision Systems, Arlington VA Date: Tue, 15 Aug 89 14:33:05 +0000 > 6. The "popularity" of neural net research is a consequence of the > miserable mathematical backgrounds of computer science students (and > some professors!).... A sweeping generalization. Computer scientists aren't the only ones working on neural networks and not all computer scientists are "student hackers." I wish people would stop confusing "programming" activities with thinking and research activities. They are distinctly different. One is engineering and the other is not. There are computer scientists who approach problems as theoreticians and there are computer scientists who approach problems with ad hoc solutions in mind. However...... (time to get up on my soapbox oh boy!).... In my opinion we don't have a deep macroscopic understanding of what neural nets are capable of doing or are doing even in the simplest networks. Researchers are spending a lot of time and effort focusing on the optimization of small techniques (e.g. backpropigation) and too little time on developing formalisms for describing and understanding NNs as a whole. A deep understanding of any complex paradigm will be reached only through the efforts of many researchers, tackling the problem from different viewpoints (like multiple sculptors chipping away at a block of marble to reveal the statue hidden inside). It's fairly useless for all of these metaphorical artists to chip away at a big toe all at once, yet they must also posses the same overall goal and understanding of the problem, otherwise the final piece will not be consistant and balanced. Well you get the idea. ------------------------------ Subject: Paradigm Shift Response (sort of) From: worden@ut-emx.UUCP (worden) Organization: The University of Texas at Austin, Austin, Texas Date: Fri, 11 Aug 89 09:25:29 +0000 It seems to me that most NN folks are doing their honest best with what little we know now. (And a thousand curses on those few but vociferous money-sniffer dilettantes!!) As I understand it, our sensory and motor systems are highly structured, from the peripheral nerves to at least several cortical layer depths. Beyond that, through the association areas and into the deeper structures of the limbic system, no one really knows what the h--l is going on. So, it doesn't surprise me that most NN folks work with the structured networks. After all, there can thereby be hope that one's model will be biologically verified. And, such work is not without merit; there remains a great deal to be understood, even in the sensory/motor systems about which we know the most. My personal preference, however, is for the random type of networks. Not as sensory/motor systems, but as possible models of the deeper systems. I have a nice micrograph, from the old 1979 Scientific American special issue on the brain, that shows a tangled mass of stained brain tissue. Apparent randomness, at least, does seem to coexist with structure inside our skulls! What I would really like to see, though, if it is not too premature, is collaboration between you majority structure enthusiasts and us minority randomness aficionados, along with some A.I. folks, to seriously attempt to build a "complete" system. My thought would be to use structured NN's for sensor input/processing and low-level learning, feeding into random NN's for multisensor fusion and mid- level learning, feeding into an A.I. subsystem for high- level learning and decision-making, feeding into random NN's for multi-effector fission, feeding into structured NN's for pre-effector conditioning and effector output. By the way, if any of you have any references to recent collaborative work between structured NN and A.I. folks, I would be very interested in getting them. Please email the info to me (or to this newsgroup). Finally, I believe that all of us are lacking critical, fundamental knowlege of some kind about how our brains work and that it is this deficiency that now prevents us from building systems that behave the way we would really like (i.e., in a "truly intelligent" fashion). I just cannot buy the arguments that greater size or greater speed or greater complexity or even greater biological realism is the "answer". I do believe that part of the answer lies in building hybrid systems, but I think that there is a deeper mystery. Perhaps some cellular function that has yet to be observed and/or understood. Perhaps an interaction between neurons and glia, as suggested by that recent Scientific American article. Perhaps some phenomenon that we don't even suspect at this point... - Sue Worden Electrical and Computer Engineering University of Texas at Austin ------------------------------ Subject: Re: Paradigm Shift Response (sort of) From: Joe Keane <jk3k+@ANDREW.CMU.EDU> Organization: Mathematics, Carnegie Mellon, Pittsburgh, PA Date: 13 Aug 89 02:41:18 +0000 In article <16946@ut-emx.UUCP> worden@ut-emx.UUCP (worden) writes: >As I understand it, our sensory and motor systems are highly >structured, from the peripheral nerves to at least several >cortical layer depths. Beyond that, through the association >areas and into the deeper structures of the limbic system, >no one really knows what the h--l is going on. Not yet at least. >Apparent randomness, at least, does seem to coexist with >structure inside our skulls! If you looked at a microprocessor chip you might say the same thing. I don't think biological neural nets are as structured as silicon chips, or we might be looking for `grandmother cells'. But i don't think they're completely random either. It's up to NN people and neurobiologists to figure out which structures are useful. ------------------------------ End of Neurons Digest *********************