neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (08/14/90)
Neuron Digest Monday, 13 Aug 1990 Volume 6 : Issue 47 Today's Topics: Summary (long): pattern recognition comparisons 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: Summary (long): pattern recognition comparisons From: Richard Fozzard <fozzard@boulder.Colorado.EDU> Date: Thu, 26 Jul 90 10:34:46 -0600 [[ Editor's Note: I've deleted extraneous headers and inserted "<***********>" between individual entries. -PM ]] Here are the responses I got for my question regarding comparisons of connectionist methods with traditional pattern recognition techniques. I believe Mike Mozer (among others) puts it best: "Neural net algorithms just let you do a lot of the same things that traditional statistical algorithms allow you to do, but they are more accessible to many people (and perhaps easier to use)." Read on for the detailed responses. (Note: this does not include anything posted to comp.ai.neural-nets, to save on bandwidth) rich ======================================================================== Richard Fozzard "Serendipity empowers" Univ of Colorado/CIRES/NOAA R/E/FS 325 Broadway, Boulder, CO 80303 fozzard@boulder.colorado.edu (303)497-6011 or 444-3168 <***********> Date: Tue, 17 Jul 90 13:53:55 -0500 From: honavar@cs.wisc.edu (Vasant Honavar) Received: by goat.cs.wisc.edu; Tue, 17 Jul 90 13:53:55 -0500 Subject: pattern recognition with nn Honavar, V. & Uhr, L. (1989). Generation, Local Receptive Fields, and Global Convergence Improve Perceptual Learning in Connectionist Networks, In: Proceedings of the 1989 International Joint Conference on Artificial Intelligence, San Mateo, CA: Morgan Kaufmann. Honavar, V. & Uhr, L. (1989). Brain-Structured Connectionist Networks that Perceive and Learn, Connection Science: Journal of Neural Computing, Artificial Intelligence and Cognitive Research, 1 139-159. Le Cun, Y. et al. (1990). Handwritten Digit Recognition With a Backpropagation Network, In: Neural Information Processing Systems 2, D. S. Touretzky (ed.), San Mateo, CA: 1990. Rogers, D. (1990). Predicting Weather Using a Genetic Memory: A Combination of Kanerva's Sparse Distributed Memory With Holland's Genetic Algorithms, In: Neural Information Processing Systems 2, D. S. Touretzky (ed.), San Mateo, CA: 1990. <***********> Date: Tue, 17 Jul 90 16:39:12 EDT From: perry@seismo.CSS.GOV (John Perry) Richard, It depends on which neural network you are using, and the underlying complexity in seperating pattern classes. We at ENSCO have developed a neural network architecture that shows far superior performance over traditional algorithms. Mail me if you are interested. John L. Perry ENSCO, Inc. 5400 Port Royal Road Springfiedl, Virginia 22151 (Springfield) 703-321-9000 email: perry@dewey.css.gov, perry@beno.css.gov <***********> Date: 17 Jul 90 14:36:00 MDT From: "Dave Shaw" <shaw_d@clipr.colorado.edu> Subject: RE: Networks for pattern recognition problems? Rich- our experience with the solar data is still inconclusive, but would seem to indicate that neural nets have exhibit no distinct advantage over more traditional techniques, in terms of 'best' performance figures. The reason appears to be that although the task is understood to be non-linear, (which should presumably lead to better performance by non-linear systems such as networks), there is not enough data at the critical points to define the boundaries of the decision surface. This would seem to be a difficulty that all recognition problems must deal with. Dave <***********> Date: Tue, 17 Jul 90 14:02:44 PDT From: kortge@galadriel.Stanford.EDU (Chris Kortge) You may know of this already, but Gorman & Sejnowski have a paper on sonar return classification in Neural Networks Vol. 1, #1, pg 75, where a net did better than nearest neighbor, and comparable to a person. I would be very interested in obtaining your list of "better-than- conventional-methods" papers, if possible (maybe the whole connectionists list would, for that matter). Thanks-- Chris Kortge kortge@psych.stanford.edu <***********> Date: Tue, 17 Jul 90 17:06:14 CDT From: galem@mcc.com (Gale Martin) I do handwriting recognition with backprop nets and have anecdotal evidence that the nets do better than the systems developed by some of the research groups we work with. The problem with such comparisons is that the success of the recognition systems depend on the expertise of the developers. There will never be a definitive study. However, I've come to believe that such accuracy comparisons miss the point. Traditional recognition technologies usually involve alot of hand-crafting (e.g., selecting features) that you can avoid by using backprop nets. For example, I can feed a net with "close to" raw inputs and the net learns to segment it into characters, extract features, and classify the characters. You may be able to do this with traditional techniques, but it will take alot longer. Extending the work to different character sets becomes prohibitive; whereas it is a simple task with a net. Gale Martin MCC Austin, TX <***********> From: "Ted Stockwell" <ted@aps1.spa.umn.edu> Subject: Re: Networks for pattern recognition problems? Date: Tue, 17 Jul 90 17:12:33 CDT > Do you know of any references to work done using connectionist (neural) > networks for pattern recognition problems? I particularly am interested > in problems where the network was shown to outperform traditional algorithms. > > I am working on a presentation to NOAA (National Oceanic and Atmospheric > Admin.) management that partially involves pattern recognition > and am trying to argue against the statement: > "...results thus far [w/ networks] have not been notably more > impressive than with more traditional pattern recognition techniques". > This may not be quite what you're looking for, but here are a few suggestions: 1) Pose the question to salespeople who sell neural network software. They probably have faced the question before. 2) One advantage is that the network chacterizes the classes for you. Instead of spending days/weeks/months developing statistical models you can get a reasonable classifier by just handing the training data to the network and let it run overnight. It does the work for you so development costs should be much lower. 3) Networks seem to be more often compared to humans than to other software techniques. I don't have the referrences with me, but I recall that someone (Sejnowski?) developed a classifier for sonar signals that performed slightly better than human experts (which *is* the "traditional pattern recognition technique"). Ted Stockwell U of MN, Dept. of Astronomy ted@aps1.spa.umn.edu Automated Plate Scanner Project <***********> From: mariah!yak@tucson.sie.arizona.edu Date: Tue, 17 Jul 90 14:58:40 -0700 Dear Dr. Fozzrad, I read your email message on a call for pattern recog. problems for which NN's are known to outperform traditional methods. I've worked in statistics and pattern recognition for some while. Have a fair number of publications. I've been reading th neural net literature and I'd be quite surprised if you get convincing replies in the affirmative, to your quest. My opinion is that stuff even from the '60's and '70's, such as the books by Duda and Hart, Gonzales and Fu, implemented on standard computers, are still much more effective than methodology I've come across using NN algorithms, which are mathematically much more rstrictive. In brief, if you hear of good solid instances favorable to NN's, please let me know. Sincerely, Sid Yakowitz Professor <***********> Date: Tue, 17 Jul 90 21:18:33 EDT From: John.Hampshire@SPEECH2.CS.CMU.EDU Rich, Go talk with Smolensky out there in Boulder. He should be able to give you a bunch of refs. See also works by Lippmann over the past two years. Barak Pearlmutter and I are working on a paper that will appear in the proceedings of the 1990 Connectionist Models Summer School which shows that certain classes of MLP classifiers yield (optimal) Bayesian classification performance on stochastic patterns. This beats traditional linear classifiers... There are a bunch of results in many fields showing that non-linear classifiers out perform more traditional ones. The guys at NOAA aren't up on the literature. One last reference --- check the last few years of NIPS and (to a lesser extent) IJCNN proceedings NIPS = Advances in Neural Information Processing Systems Dave Touretzky ed., Morgan Kaufmann publishers ICJNN = Proceedings of the International Joint Conference on Neural Networks, IEEE Press John <***********> Date: Tue, 17 Jul 90 20:51:28 MDT From: Michael C. Mozer <mozer@neuron> Subject: Re: Help for a NOAA connectionist "primer" Your boss is basically correct. Neural net algorithms just let you do a lot of the same things that traditional statistical algorithms allow you to do, but they are more accessible to many people (and perhaps easier to use). There is a growing set of examples where neural nets beat out conventional algorithms, but nothing terribly impressive. And it's difficult to tell in these examples whether the conventional methods were applied appropriately (or the NN algorithm in cases where NNs lose to conventional methods for that matter). Mike <***********> Date: Wed, 18 Jul 90 00:40:21 EDT From: burrow@grad1.cis.upenn.edu (Tom Burrow) Subject: procedural vs connectionist p.r. Status: R Sorry, this isn't much of a contribution -- mostly a request for your replies. If you are not going to repost them via the connectionist mailing list, could you mail them to me? No, for my mini-contribution: Yan Lecun, et al's work as seen in NIPS 90 on segmented character recognition is fairly impressive, and they claim that their results are state of the art. Tom Burrow <***********> Date: Wed, 18 Jul 90 09:32:08 CDT From: jsaxon@cs.tamu.edu (James B Saxon) Subject: Re: Networks for pattern recognition problems? In article <23586@boulder.Colorado.EDU> you write: >Do you know of any references to work done using connectionist (neural) >networks for pattern recognition problems? I particularly am interested >in problems where the network was shown to outperform traditional algorithms. > >I am working on a presentation to NOAA (National Oceanic and Atmospheric >Admin.) management that partially involves pattern recognition >and am trying to argue against the statement: >"...results thus far [w/ networks] have not been notably more >impressive than with more traditional pattern recognition techniques". > >I have always felt that pattern recognition is one of the strengths of >connectionist network approaches over other techniques and would like >some references to back this up. Well, aside from the question of ultimate generality to which the answer is "OF COURSE there are references to neural network pattern recognition systems. The world is completely full of them!" Anyway, maybe you'd better do some more research. Here's a couple off the top of my head: Kohonen is really hot in the area, he's been doing it for at least ten years. Everybody refers to some aspect of his work. I also suggest picking up a copy of the IJCNN '90 San Diego, all 18lbs of it. (International Joint Conference on Neural Networks) But for a preview: I happened to sit in on just the sort of presentation you would have liked to hear. His title was "Meterological Classification of Satellite Imagery Using Neural Network Data Fusion" Oh Boy!!! Big title! Oh, it's by Ira G. Smotroff, Timothy P. Howells, and Steven Lehar. >From the MITRE Corporation (MITRE-Bedford Neural Network Research Group) Bedford, MA 01730. Well the presentation wasn't to hot, he sort of hand waved over the "classification" of his meterological data though he didn't describe what we were looking at. The idea was that the system was supposed to take heterogeneous sensor data (I hope you know these: GOES--IR and visual, PROFS database--wind profilers, barometers, solarometers, thermometers, etc) and combine them. Cool huh. If they had actually done this, I imagine the results would have been pretty good. It seems though that they merely used an IR image and a visual image and combined only these two. Their pattern recognition involved typical modeling of the retina which sort of acts as a band pass filter with orientations, thus it detects edges. Anyway, their claim was the following: "The experiments described showed reasonably good classification performance. There was no attempt to determine optimal performance by adding hidden units [Hey, if it did it without hidden units, it's doing rather well.], altering learning parameters, etc., because we are currently implementing self-scaling learning algorithms which will determine many of those issues automatically. [Which reminds me, Lehar works with Grossberg at Boston University. He's big on pattern recognition too, both analog and digital. Check out Adaptive Resonance Theory, or ART, ART2, ART3.]..." Anyway it looks like a first shot and they went minimal. There's lots they could add to make it work rather well. In terms of performance, I'd just like to make one of those comments... From what I saw at the conference, neural networks will outperform traditional techniques in this sort of area. The conference was brimming over with successful implementations. Anyway... Enough rambling, from a guy who should be writing his thesis right now... Good luck on your presentation! Oh, I think it automacally puts my signature on..... Did it? /--------------------------------------------\ James Bennett Saxon | "I aught to join the club and beat you | Visualization Laboratory | over the head with it." -- Groucho Marx | Texas A&M University <---------------------------------------------/ jsaxon@cssun.tamu.edu <***********> Date: Wed, 18 Jul 90 10:38:16 EDT From: Henri Arsenault <arseno@phy.ulaval.ca> Subject: papers on pattern recognition In response to your request about papers on neural nets in pattern recognition, there is a good review in IEEE transactions on neura l networks, vol. 1, p. 28. "Survey of neural network technology for automatic target recognition", by M. W. Roth. The paper has many references. arseno@phy.ulaval.ca <***********> Date: Wed, 18 Jul 90 09:07:50 PDT From: d38987%proteus.pnl.gov@pnlg.pnl.gov Subject: NN and Pattern Recognition Richard, We have done some work in this area, as have many other people. I suggest you call Roger Barga at (509)375-2802 and talk to him, or send him mail at: d3c409%calypso@pnlg.pnl.gov Good luck, Ron Melton Pacific Northwest Laboratory Richland, WA 99352 <***********> To: mozer@neuron Subject: Re: Help for a NOAA connectionist "primer" Status: R mike, thanks for the input - it seems a cogent summary of the (many) responses I've been getting. However, it seems just about noone has really attempted a one-to-one sort of comparison using traditional pattern recognition benchmarks. Just about everything I hear and read is anecdotal. Would it be fair to say that "neural nets" are more accessible, simply because there is such a plethora of 'sexy' user-friendly packages for sale? Or is back-prop (for example) truly a more flexible and widely-applicable algorithm than other statistical methods with uglier-sounding names? If not, it seems to me that most connectionists should be having a bit of a mid-life crisis about now. rich <***********> Date: Wed, 18 Jul 90 11:16:22 MDT From: Michael C. Mozer <mozer@neuron> Subject: Re: Help for a NOAA connectionist "primer" I think NNs are more accessible because the mathematics is so straightforward, and the methods work pretty well even if you don't know what you're doing (as opposed to many statistical techniques that require some expertise to use correctly). For me, the win of NNs is as a paradigm for modeling human cognition. Whether the NN learning algorithms existed previously in other fields is irrelevant. What is truly novel is that we're bringing these numerical and statistical techniques to the study of human cognition. Also, connectionists (at least the cog sci oriented ones) are far more concerned with representation -- a critical factor, one that has been much studied by psychologists but not by statisticians. Mike <***********> From: Ron Cole <cole@cse.ogi.edu> Subject: Re: Networks for pattern recognition problems? Call Les Atlas at U Washington. He has an article coming out in IEEE Proceedings August comparing NNs and CART on 3 realworld problems. Ron Les Atlas: 206 685 1315 <***********> Date: Wed, 18 Jul 90 10:59:24 PDT From: bimal@jupiter.risc.com (Bimal Mathur) Subject: pattern recognition The net result of experiments done by us in pattern classification for two dimensinal data i.e. image to features, classify features using NN, is that there is no significant improvement in performance of the overall system. -bimal mathur - Rockwell Int <***********> Date: Wed, 18 Jul 90 14:10:28 EDT From: Chip Bachmann <PH706008@brownvm.brown.edu> Subject: Re: Networks for pattern recognition problems? An example of research directly comparing neural networks with traditional statistical methods can be found in: R. A. Cole, Y. K. Muthusamy, and L. Atlas, "Speaker-Independent Vowel Recognition: Comparison of Backpropagation and Trained Classification Trees", in Proceedings of the Twenty-Third Annual Hawaii International Conference on System Sciences, Kailua-Kona, Hawaii, January 2-5, 1990, Vol. 1, pp. 132-141. The neural network achieves better results than the CART algorithm, in this case for a twelve-class vowel recognition task. The data was extracted from the TIMIT database, and a variety of different encoding schemes was employed. Tangentially, I thought that I would enquire if you know of any postdoctoral or other research positions available at NOAA, CIRES, or U. of Colorado. I completed my Ph.D. in physics at Brown University under Leon Cooper (Nobel laureate, 1972) this past May; my undergraduate degree was from Princeton University and was also in physics. My dissertation research was carried out as part of an interdisciplinary team in the Center for Neural Science here at Brown. The primary focus of my dissertation was the development of an alternative backward propagation algorithm which incorporates a gain modification procedure. I also investigated the feature extraction and generalization of backward propagation for a speech database of stop-consonants developed here in our laboratory at Brown. In addition, I discussed hybrid network architectures and, in particular, in a high-dimensional, multi-class vowel recognition problem (namely with the data which Cole et. al. used in the paper which I mentioned above), demonstrated an approach using smaller sub-networks to partition the data. Such approaches offer a means of dealing with the "curse of dimensionality." If there are any openings that I might apply for, I would be happy to forward my resume and any supporting materials that you might require. Charles M. Bachmann Box 1843 Physics Department & Center for Neural Science Brown University Providence, R.I. 02912 e-mail: ph706008 at brownvm <***********> Date: Wed, 18 Jul 90 16:05:39 EDT From: Charles Wilson x2080 <wilson@magi.ncsl.nist.gov> Organization: National Institute of Standards and Technology formerly National Bureau of Standards Subject: character recognition We have shown on character recognition problem that neural networks are as good in accuracy as traditional methods but much faster (on a parallel computer), much easier to prpgram ( a few hundred lines of parallel fortran) and less brittle. see C. L. Wilson, R. A. Wilkinson, and M. D. Garris, "self-organizing Neural Network Character Recognition on a Massively Parallel Computer", Proc. of the IJCNN, vol 2, pp. 325-329, June 1990. <***********> Date: 18 Jul 90 16:22:00 MDT From: "Dave Shaw" <shaw_d@clipr.colorado.edu> Subject: RE: Networks for pattern recognition problems? To date we have compared the expert system originally built for the task with many configurations of neural nets (based on your work), multiple linear regression equations, discriminant analysis, many types of nearest neighbor systems, and some work on automatic decision tree generation algorithms. Performance in measured both in the ROC P sub a (which turns out to be only a moderate indicator of performance, due to the unequal n's in the two distributions), and maximum percent correct, given the optimal bias setting. All systems have been trained and tested on the same sets of training and test data. As I indicated before, the story isn't completely in yet, but it is very hard to show significant differences between any of these systems on the solar flare task. Dave <***********> Date: Wed, 18 Jul 90 16:45:41 EST From: George Kaczowka <uvm-gen!idx!gsk@uunet.UU.NET> Subject: Networks for pattern recognition problems? Rich -- I don't know if this helps, but a company in Providence RI called NESTOR has put together a couple of products.. come of which have been customized systems for customers solving pattern recognition problems.. One I remember was regarding bond trading in the financial world.. I seem to remenber that the model outperformed the "experts" by at least 10-15%, and that this was used (and is as far as I know) by some on wall street. I know that they have been in the insurance field for claim analysys as well as physical pattern recognition.. They were founded by a phd out of Brown University, and I am sure that you caould obtain reference works from them.. I understand that they are involved in a few military pattern recognition systems for fighters as well.. Good luck.. I was interested in their work some time ago, but have been off on other topics for over a year.. -- George -- ------------------------------------------------------------ - George Kaczowka IDX Corp Marlboro, MA - gsk@idx.UUCP - ------------------------------------------------------------ <***********> Date: Fri, 20 Jul 90 08:48:46 EST From: marwan@ee.su.oz.AU (Marwan Jabri) Subject: pattern recognition We have been working on the application of neural nets to the pattern r recognition of ECG signals (medical). I will be happy in mailing you some of our very good results that are better of what has been achieved using conventional techniques. Is this the sort of things you are looking for? what media you want? Marwan Jabri ------------------------------------------------------------------- Marwan Jabri, PhD Email: marwan@ee.su.oz.au Systems Engineering and Design Automation Tel: (+61-2) 692-2240 Laboratory (SEDAL) Fax: (+61-2) 692-3847 Sydney University Electrical Engineering NSW 2006 Australia <***********> Date: Fri, 20 Jul 90 10:10:26 CST From: PP219113@tecmtyvm.mty.itesm.mx Subject: Re: Networks for pattern recognition problems? Organization: Instituto Tecnologico y de Estudios Superiores de Monterrey hi, David J. Burr (in 'Experiments on NN Recognition of Spoken and Written Text, suggests that NN and Nearest neighbor classification performs at near the same level of accuracy, IEEE Trans on ASSP, vol 36, #7,pp1162-68, july 88) My own experience with character recognition using neural nets actually suggests that NN have better performance than nearest neighbor and hierarchical clustering, (I suggest to talk to Prof. Kelvin Wagner, ECE, UC-Boulder) See also, "Survery of Neural Net Tech for Automatic Target Recognition" in Trans in Neural Net, March 90, pp 28 by M.W. Roth. jose luis contreras-vidal <***********> Date: 20 Jul 90 15:32:00 MDT From: "Dave Shaw" <shaw_d@clipr.colorado.edu> Subject: RE: Networks for pattern recognition problems? Rich- the network configurations we have used are all single hidden layer of varying size (except for 1 network with no hidden layer). Hidden layer size has been varied from 1 to 30 units. Input layer=17 units, output layer=1 unit. All activation functions sigmoidal. As I indicated before, there was essentially no difference between any of the networks. We are moving towards a paper (one at least) and this work will likely be included as part of my dissertation as well. Dave ------------------------------ End of Neuron Digest [Volume 6 Issue 47] ****************************************