« iepriekšējāTurpināt »
when no examples of both faults occur- of Southern California (USC) uses the Hong Kong University of Science ring together are included in the train- Hebbian type of correlation for mem- and Technology, there is work on neting set. The main conclusion is that the ory storage. However, by using works whose processing elements are spectral representation provides the Householder encoding, the memory quadratic functions and work on mapbest recognition performance and that capacity of the BAM is greatly improved ping multilayer attributed graphs onto the primary discriminatory informa- over the conventional BAM. Also at a neocognitron network. tion for distinguishing the two types of faults (impeller unbalance and cracked
Table 1. Breakdown of Authors and impeller) is contained in the first three
Delegates by Country harmonics. This work is relevant to the Navy's research program on helicopter
Authors Delegates gearbox fault diagnosis.
6 The neural network research in China Brazil
2 is also spread out in many universities. Bulgaria
1 The work is very much application Canada
14 oriented, as can be seen by the list of Chile
1 topics and locations in Table 3.
16 Of particular interest to the Navy is Czechoslovakia
1 the work of Jungang Xu, Zhong Wang,
5 and Youan Ke at the Beijing Institute France
11 of Technology on optimum frequency Germany
16 selection for radar target classification Greece
1 by a neural network. The radar cross Hong Kong
11 sections at multiple frequencies are used India
4 as the inputs to a backpropagation neural Indonesia
1 network for radar target classification. Ireland
1 The frequencies corresponding to the Israel
2 input nodes that have maximum sensi- Italy
11 tivities are selected as optimum fre- Japan
125 quencies needed for classification. The Korea
19 method is applied successfully in clas- Mexico
1 sifying simple radar targets in an Malaysia
1 Anechoic Chamber. Obviously this Netherlands
4 result has significance in automatic radar New Zealand
2 target classifications for the Navy. Norway
103 HONG KONG South Africa
2 Research is concentrated in the three Sweden
1 big universities in Hong Kong. At the Switzerland
2 Hong Kong University of Science and Taiwan
11 Technology, there is work on the auto- Thailand
1 matic determination of multilayer feed- Turkey
1 forward network size for supervised United Arab Emirate
1 learning and work on using the United Kingdom
19 Householder encoding algorithm for U.S.A.
91 discrete bidirectional associative U.S.S.R.
1 memory (BAM). The original BAM as Yugoslavia
2 proposed by Bart Kosko of the University
Research Topics and Locations in Australia
On-line identification of nonlinear systems
Univ. of Western Australia
At the Chinese University of Hong Basis Function (RBF) classifier of traditionally they have been very strong Kong, the neural network research Thomas Poggio.
in learning automatons, but only a small efforts are spearheaded by Professor At City Polytechnic of Hong Kong, number of papers were presented at Lai-Wan Chan, who is a graduate of a new Hopfield type of training algo- this conference. the Imperial College in London. They rithm has been developed whereby the At the Indian Institute of Science, a have developed a novel learning algo- orthogonally coded memories are great deal of work is devoted to the rithm for a recurrent backpropagation obtained iteratively. Also a structured study of bidirectional associative network and a system for detecting three- backpropagation network has been memories (BAM). The extended hyperdimensional (3-D) motion from a developed that adapts to the problem. cube, as well as the INMOS transputersequence of image frames. At the Clearly these are initial efforts in apply
Clearly these are initial efforts in apply- based computer system, is used to Chinese University, BAM is studied ing neural network technology. simulate the parallelism in BAM. The from the viewpoint of match filtering.
Fokker-Planck equation is also used to Sufficient conditions are found for INDIA
study the dynamics of associative stability and attractivity for BAM net
memories. The Hopfield-based model works. Also there is work on neural Research appears to be centered at for distributed representation of objects networks that learn the decision bound- the Indian Institute of Science and the and for use as content addressable aries with nonlinear clustering, similar Center for Artificial Intelligence and memory is also investigated. There is to the Reduced Coulomb Energy (RCE) Robotics. It seems that there should be also research work on genetic algomodel of Leon Cooper and the Radial a lot more work going on in India because rithms and learning automata for pat
of particular interest to the Navy is number of researchers involved in neural Top-down views of three different airthe work of P.Y. Mundkur and networks is close to a thousand. The craft (DC-10, Phantom, and MiG21) at U.B. Desai at the Indian Institute of papers, close to a hundred, presented different azimuth and roll angles are Technology on automatic target recog- at this conference are but a small por- used for training the network. The trainnition. A sequence of images taken tion of the neural network work that is ing set consists of 216 images (72 images from high altitude reconnaissance flights going on in Japan. Table4 is a list of the per aircraft at 10° roll and azimuth is used as input to the network. The topics of research at various universi- angles). Two multilayer feedforward network consists of two cascaded ties and industry.
neural nets are used, one for classifying modules, a multilayer perceptron net
the type of aircraft and another for work followed by a modified Maxnet, KOREA
estimating the orientation angle. Out to provide translation-invariant recog
of the three aircraft types, correct clasnition of the targets in clutter. The Most of the neural network research sification approaches the 98.6% level, network is partitioned into subnets and is centered at the Korea Advanced and the accuracy of estimating the oria backpropagation algorithm is used to Institute of Science and Technology, entation is at 89.4%. Of course, these train the subnets. Preliminary results although there are a few efforts at the performance levels are extremely good show promise, but a lot more research other universities and industry. Of for such a small number of different needs to be done yet.
particular interest to the Navy is the aircraft types. This work has implica
work of Dae-Young Yim, Sung-Il Chien, tions in military automatic target recogJAPAN
and Hyun Son at Kyungpook National nition as well as in commercial avia
University on multiclass 3-D identifi- tion. Table 5 is a listing of the location Japan clearly is the most technolog- cation and orientation estimation using of different neural network research in ically advanced country in Asia. The multilayer feedforward neural networks. Korea.
Hebbian type associative memory
Receptive field network for kanji recognition
lving the four color mapping prob em
Research Topics and Locations in Korea
Iterative autoassociative memory
yota Tech. Inst.
Kyungpook National Univ.