Applications of Artificial Neural Networks Database Maintained by Wesley R. Elsberry P.O. Box 1187 Richland, WA 99352 or Release 1.0, January 31, 1992 Copyright 1992 by Wesley R. Elsberry This document is released for redistribution on free-access electronic information sources. Other publication requires the permission of the author. The following entries give a brief description of an application of artificial neural network (ANN) technology. Each entry indicates the type of model used in the application, where it was developed, and a reference for more detailed study. I solicit contributions to this database. Please format them like the entries given below and send via email to , via NetMail to "Wesley R. Elsberry, 1:347/303", via EchoMail as a message on the NEURAL_NET Echo, or via SnailMail on MS-DOS floppy to the address given above. Application : Adaptive noise cancellation in fetal electrocardiography Model : ADALINE Hardware/ Software: ? Organization : Stanford University Notes : Use of leads from upper chest and abdomen of mother allow for noise cancellation, thus "revealing" the fetal cardiac waveforms. Reference : Widrow, B., & R. Winter. 1988. Neural nets for adaptive filtering and adaptive pattern recognition. IEEE Computer (March 1988):25-39. Application : Channel equalization Model : ADALINE Hardware/ Software: ? Organization : Stanford University Notes : Adaptive equalizer uses incoming signals as training response to develop channel inverse, thus compensating for irregularities. "Any least-squares algorithm can adapt the weights, but the telecommunications industry uses the LMS algorithm almost exclusively." Reference : Widrow, B., & R. Winter. 1988. Neural nets for adaptive filtering and adaptive pattern recognition. IEEE Computer (March 1988):25-39. Application : Learning Mortgage Underwriting Decisions Model : Restricted Coulomb Energy in multiple subnetworks (MNNLS) Hardware/ Software: ? Organization : Nestor, Inc. Notes : An ANN here was trained on mortgage applications, looking at the approve / disapprove decision. The network was trained to be conservative. The test cases were ordered by level of difficulty. At th simpler end of the scale, the network agreed with the underwriter 96% of the time; at the difficult end, 82% of the time. Differences here stem mainly from a lack of consistency in the underwriter's appraisal. Reference : Collins, E., S. Ghosh, & C. Scofield. 1988. An application of a multiple neural network learning system to emulation of mortgage underwriting judgements. Proc. IEEE ICNN 1988 2:459-466. Application : Diagnosis of skin disease Model : Back-propagation (BP) Hardware/ Software: Software Organization : University of Texas Southwestern Medical Center Notes : Diagnosis of papulosquamous disease using expert diagnosis as training signal for BP. Performed at about 67% correct diagnosis, which is pretty good given that the field is dermatology. Reference : Yoon, Y.O., L.L. Peterson, & P.R. Bergstrasser. 1988. A dermatology expert system using connectionist network. Presented at IEEE ICNN 1988. Application : Diagnosis of low back pain Model : Multi-layer Perceptron Hardware/ Software: Software Organization : Royal Signals and Radar Establishment, Frenchay Hospital, Western Infirmary (all groups are in the UK) Notes : Multi-layer perceptrons were trained on back pain diagnoses. In tests against three different groups of doctors and a fuzzy logic system, the only group of diagnoses which the MLP performed less well than the others was simple low back pain. MLP outperformed other groups for Spinal Pathology test cases, which were the most critical. Reference : Bounds, D.G., P.J. Lloyd, B. Mathew, & G. Waddell. 1988. A multi layer Perceptron network for the diagnosis of low back pain. Proc. IEEE ICNN 1988 2:481-489. Application : Stock market prediction Model : Modified BP Hardware/ Software: Software Organization : Fujitsu & Nikko Securities Notes : The network system showed a profit in tests. Real-life investment testing to begin soon. Reference : Kimoto, T., & K. Asakawa. 1990. Stock market prediction with modular neural networks. Proc. IJCNN 1990 (San Diego) 1:1. Application : Electron beam lithography Model : BP Hardware/ Software: Software training/hardware product Organization : AT&T Bell Labs Notes : Corrections for proximity effect in electron beam lithography can be computed using an iterative method, but it's slow. Software ANN performed to within 0.5% of iterative results, and was 37 times faster. Hardware version built using weights from software version was over 1000 times faster. Reference : Frye, R.C, E.A. Rietman, & K.D. Cummings. 1990. Computation of proximity effect corrections in electron beam lithography by a neural network. Proc. IJCNN 1990 (San Diego) 1:7-14. Application : Fault diagnosis Model : Neural-Logic Network Hardware/ Software: Software Organization : Delco Electronics Division Notes : Diagnosing faults in an line-replaceable unit (LRU) inertial navigation system (INS) avionics system, the network gave a diagnosis in 64% of cases (47.6% complete, 16.4% partial). Seen as useful in training new technicians. Reference : Tan, A.H., Q.Pan, & H.H. Teh. 1990. INSIDE: Aneuronet based hardware fault diagnostic system. Proc. IJCNN 1990 (San Diego) 1:63-68. Application : Non-invasive Structural analysis Model : BP Hardware/ Software: Software Organization : Ministry of Transportation at Ontario and the Communications Research Laboratory at McMaster University Notes : Assessing condition of asphalt-covered bridge decks using impulse radar waveforms. The network is given reduced data (derived using principal components analysis) and produces classifications which are 95 to 100% accurate. Reference : Vrckovnik, G., T. Chung, & C.R. Carter. 1990. Classifying impulse radar waveforms using principal components analysis and neural networks. Proc. IJCNN 1990 (San Diego) 1:69-74. Application : Air combat maneuver selection Model : BP Hardware/ Software: Software Organization : USAF Human Resources Lab Notes : 38 possible maneuvers, 40 scenarios. Production system agreed with pilots in 10 scenarios, ANN agreed in 27. ANN over 2.5x better than the production system at selecting a maneuver. Reference : McMahon, D.C. 1990. A neural network trained to select aircraft maneuvers during air combat: a comparison of network and rule based performance. Proc. IJCNN 1990 (San Diego) 1:107-112. Application : Seismic first-break picking Model : BP Hardware/ Software: Software Organization : Amoco Notes : Seismic data analysis involves the removal of spurious reflected waves in the, which requires identification of the first break signal. NN produced 95% accurate results in about a minute for a run of 53 records (81 traces per record). The accuracy exceeds that of current computational methods, the speed shows an eight-fold improvement over manual picking. Reference : Veezhinathan, J., & D. Wagner. 1990. A neural network approach to first break picking. Proc. IJCNN 1990 (San Diego) 1:235-240. Application : Loan scoring Model : BP Hardware/ Software: Software w/hardware accelerator Organization : HNC Notes : Financial service tested ANN for loan scoring using a database of 17,000 loans. 10,000 were used for training, the remaining 7,000 for testing. "The results imply that use of neural network technology could increase overall profitability up to 27% over the existing approval method, in which the company staff uses a computer-based credit scoring algorithm." Reference : Flyer: "HNC in Financial Services" Application : Inspection and quality control Model : BP Hardware/ Software: Software w/hardware accelerator Organization : HNC Notes : Video image is processed using FFT, then presented to back-propagation network for classification. This was demonstrated at 1988 ICNN for use in a product inspection system, where the product was a liquid-filled bottle. Faults recognized by the system were "fill too low," "fill too high," "label misaligned," and "top not properly in place." Reference : Flyer: "HNC in Machine Vision" Application : Bond-rating Model : BP Hardware/ Software: Software Organization : University of California (Berkeley) Notes : ANN compared to multiple linear regression technique for assigning a rating to a bond. The ANN consistently outperformed the MLR technique (80% compared to 63% on training data; 82% compared to 65% on test data). Dutta and Shakhar noted that the MLR technique sometimes gave a wildly inaccurate rating when wrong; the ANN never was off by more than one rating level. Reference : Caudill, M. 1990. Using neural networks: making an expert network. AI Expert (July):41-45. Application : Refinery process control Model : BP Hardware/ Software: Software Organization : Texaco Notes : ANN given 1,440 example training set, then set up to control 17 hour batch debutanizer process. Results indicate that the ANN is in control about 80% of the time, more if the batch is unstable. Reference : Kestelyn, J. 1990. Applications watch. AI Expert (June):71. Application : Job scheduling Model : Hopfield Hardware/ Software: Software Organization : NASA Notes : Job scheduling for the resources of the Hubble Space Telescope uses a Hopfield network in conjunction with other tools. This allows for multiple "good" candidate schedules to be generated. Reference : Kestelyn, J. 1990. Applications watch. AI Expert (June):71. __________________________________________________________________ The author will almost always allow printing for educational purposes without fee or royalty, please contact Wesley R. Elsberry for details.