Lapas attēli
PDF
ePub

when no examples of both faults occurring together are included in the training set. The main conclusion is that the spectral representation provides the best recognition performance and that the primary discriminatory information for distinguishing the two types of faults (impeller unbalance and cracked impeller) is contained in the first three harmonics. This work is relevant to the Navy's research program on helicopter gearbox fault diagnosis.

CHINA

The neural network research in China is also spread out in many universities. The work is very much application oriented, as can be seen by the list of topics and locations in Table 3.

Of particular interest to the Navy is the work of Jungang Xu, Zhong Wang, and Youan Ke at the Beijing Institute of Technology on optimum frequency selection for radar target classification by a neural network. The radar cross sections at multiple frequencies are used as the inputs to a backpropagation neural network for radar target classification. The frequencies corresponding to the input nodes that have maximum sensitivities are selected as optimum frequencies needed for classification. The method is applied successfully in classifying simple radar targets in an Anechoic Chamber. Obviously this result has significance in automatic radar target classifications for the Navy.

HONG KONG

Research is concentrated in the three big universities in Hong Kong. At the Hong Kong University of Science and Technology, there is work on the automatic determination of multilayer feedforward network size for supervised learning and work on using the Householder encoding algorithm for discrete bidirectional associative memory (BAM). The original BAM as proposed by Bart Kosko of the University

[blocks in formation]

Table 2. Research Topics and Locations in Australia

Topic

On-line identification of nonlinear systems

Time series analysis

Hybrid systems

Higher order neural nets

Rotating machine fault diagnosis

Learning in feedforward networks

Texture segmentation

Dynamic channel assignment

Classification of intracardiac electrocardiograms

Probabilistic neural network

Reinforcement learning

Object recognition for robots

Kohonen algorithm on transputers
Information retrieval

Visual reconstruction networks
Adaptive quadratic neural net
Autoassociative network

Network of exponential neurons
Robust networks

Self-organizing network for object recognition

Genetic algorithms

Unsupervised learning for neural trees Kolmogorov representation theorem

Location

Univ. of Western Australia

Univ. of Melbourne

Swinburne Inst. of Tech.

Telecom Research Labs

Royal Melbourne Inst. of Tech.
Univ. of Queensland
Monash Univ.

Univ. of Melbourne
Univ. of Sydney

Univ. of Western Australia
Univ. of Western Australia
BHP Research, Melbourne Lab
Univ. of Western Australia
Univ. of New South Wales
Telecom Research Labs
Univ. of Western Australia
Griffith Univ.

Queensland Univ. of Tech.
Univ. of New South Wales
Univ. of Melbourne

Univ. of Western Australia
Telecom Research Labs
Telecom Research Labs

Basis Function (RBF) classifier of traditionally they have been very strong
Thomas Poggio.
in learning automatons, but only a small
number of papers were presented at
this conference.

At City Polytechnic of Hong Kong, a new Hopfield type of training algorithm has been developed whereby the orthogonally coded memories are obtained iteratively. Also a structured backpropagation network has been developed that adapts to the problem. Clearly these are initial efforts in applying neural network technology.

At the Chinese University of Hong Kong, the neural network research efforts are spearheaded by Professor Lai-Wan Chan, who is a graduate of the Imperial College in London. They have developed a novel learning algorithm for a recurrent backpropagation network and a system for detecting threedimensional (3-D) motion from a sequence of image frames. At the Chinese University, BAM is studied from the viewpoint of match filtering. Sufficient conditions are found for INDIA stability and attractivity for BAM networks. Also there is work on neural networks that learn the decision boundaries with nonlinear clustering, similar to the Reduced Coulomb Energy (RCE) model of Leon Cooper and the Radial

Research appears to be centered at the Indian Institute of Science and the Center for Artificial Intelligence and Robotics. It seems that there should be a lot more work going on in India because

At the Indian Institute of Science, a great deal of work is devoted to the study of bidirectional associative memories (BAM). The extended hypercube, as well as the INMOS transputerbased computer system, is used to simulate the parallelism in BAM. The Fokker-Planck equation is also used to study the dynamics of associative memories. The Hopfield-based model for distributed representation of objects and for use as content addressable memory is also investigated. There is also research work on genetic algorithms and learning automata for pattern recognition.

[blocks in formation]

Associative memory

Self-improving associative network
Adaptive pattern recognition

Transform domain backpropagation algorithm
Laterally inhibitory neural network
Multilayer network with dynamic neurons
Recall in multilayer perceptron network
New neural network architecture
Radar target classification

Character recognition using holographic memory
Local minima in backpropagation error surface
Job-shop scheduling

Object oriented neural network language
Unstructured economic decision process
Adaptive predictor for nonlinear dynamics
Speaker independent syllable recognition
Absolute stability of Hopfield nets
Short time speech recognition
Bounds on the approximation capacity
Hopfield-Tank model for solving TSP
Weighted associative memory

Fault tolerance in self-organizing nets
Complexity of learning algorithms
Speaker recognition

Tsinghua Univ.

Zhejiang Univ.

Nanjing Aeronautical Inst.
Jiao Tong Univ.

Tsinghua Univ.

South China Univ. of Tech.
Inst. of Electronics

Inst. of Electronics
Beijing Inst. of Tech.

South China Univ. of Tech.
Southeast Univ.

Tsinghua Univ.

Changsha Inst. of Tech.

Inst. of Auto. Academia Sinica
Univ. of Sci. & Tech. of China
Beijing Inst. of Tech.
Beijing Univ. of Aeronautics
South China Univ.

South China Univ. of Tech.

Beijing Univ. of Post & Telecom. Zhejiang Univ.

Southeast Univ. Tsinghua Univ. Peking Univ.

number of researchers involved in neural
networks is close to a thousand. The
papers, close to a hundred, presented
at this conference are but a small por-
tion of the neural network work that is
going on in Japan. Table 4 is a list of the
going on in Japan. Table 4 is a list of the
topics of research at various universi-
topics of research at various universi-
ties and industry.

Of particular interest to the Navy is the work of P.Y. Mundkur and U.B. Desai at the Indian Institute of Technology on automatic target recognition. A sequence of images taken from high altitude reconnaissance flights is used as input to the network. The network consists of two cascaded modules, a multilayer perceptron network followed by a modified Maxnet, KOREA to provide translation-invariant recognition of the targets in clutter. The network is partitioned into subnets and a backpropagation algorithm is used to train the subnets. Preliminary results show promise, but a lot more research needs to be done yet.

JAPAN

Japan clearly is the most technologically advanced country in Asia. The

Most of the neural network research is centered at the Korea Advanced Institute of Science and Technology, although there are a few efforts at the other universities and industry. Of particular interest to the Navy is the work of Dae-Young Yim, Sung-Il Chien, and Hyun Son at Kyungpook National University on multiclass 3-D identification and orientation estimation using multilayer feedforward neural networks.

Top-down views of three different aircraft (DC-10, Phantom, and MiG21) at different azimuth and roll angles are used for training the network. The training set consists of 216 images (72 images per aircraft at 10° roll and azimuth angles). Two multilayer feedforward neural nets are used, one for classifying the type of aircraft and another for estimating the orientation angle. Out of the three aircraft types, correct classification approaches the 98.6% level, and the accuracy of estimating the orientation is at 89.4%. Of course, these performance levels are extremely good for such a small number of different aircraft types. This work has implications in military automatic target recognition as well as in commercial aviation. Table 5 is a listing of the location of different neural network research in Korea.

[blocks in formation]

Error correction learning in three-layer network
Inverted pendulum problem

Adaptive decision feedback equalization
Optimal control with recurrent network

Three-layer backpropagation for pattern recognition
Hierarchical Markov Random Field model

Finite size multilayer net for polynomials
Models of the cerebellum

Fuzziness in 3-D surface depth perception
Recognition of facial expressions
Network model for color blindness

Model for conscious and unconscious processing
Time delay discrete neural network
Multilayer network with quantizer neurons
Inverse modeling of dynamical systems

Fuzzy training algorithm for phoneme recognition
Memory based artificial neural network
Learning rule using difference approximation
Spatial inhibition and local association

Pattern classification for remote sensing
Moment invariants for katakana recognition
Road segment extraction from maps

Rule evaluating neural networks
Feature selection for recognition
Piecewise linear higher order network
Learning process of recurrent networks
Automatic determination of association units
Estimation of a posteriori probability
Asymptotic behavior of simulated annealing
Global suppression of spurious states
Automatic gray level adjustments

Multiply descent cost competitive learning
Regularization vision chips

CombNET-II for written digit classification
Rotation invariant neural pattern recognition
Hierarchical intelligent control

Feedforward network for robot motion control
Visual tracking of robot manipulator

Finite times of search in multilayer networks
Synthesizing networks for pattern recognition
Elimination of local minima in backpropagation
Neural sequential associator for prediction
Speaker independent 1000 word recognition
Car detection system using neocognitrons

Limb function discrimination using EMG signals
Feedforward control based on inverse systems
Learning scheme to improve speed of convergence
Quasi-symmetric logic networks

Ultrasonic 3-D visual sensor

Space perception model

Human hand position control learning

Keio Univ.

Nagoya Univ.

NEC Corp.

Kisarazu National Tech. College
Kisarazu National Tech. College

Kyoto Inst. of Tech.
Hosei Univ.

NTT Commun. Switching Lab
Kyushu Inst. of Tech.
Kyushu Inst. of Tech.
ATR Research Lab
Toshiba Corp.

[blocks in formation]
[blocks in formation]

Receptive field network for kanji recognition
Extension of backpropagation algorithm
Regression analysis with inverse model
Face graph method for fuzzy neural networks
Bidirectional optical neural net

Basin size in autoassociative memory

Learning algorithm based on information theory
Recursive neural system in a tree like structure
Reduction of precision in learning

Neural network based adaptive control system
Wafer scale LSIs with 1152 digital neurons
Low-bit learning algorithm for pattern recognition
Cross-coupled Hopfield network

Neuromorphic sensing and control

Controller for autonomous underwater robot

Trajectory generation for biped robot

Inverse kinematic calculations

Self-learning robot vision system

NN/II network for pattern recognition

Time warping network for phoneme recognition
Adaptive neural model reference structure

Structure detection by neural sequence associator
Neural searchlight processor

Cluster formation in random neural network

Parallel ASIC VLSI neurocomputer

Inducing algorithm for LTP in hippocampus
Temporal association in symmetric networks
Dynamics in chaotic neural networks

T-model neural network with learning ability
Capabilities of three-layer network
Parallel algorithm for simulated annealing
Solving the four color mapping problem

Solving the dynamic traveling salesman problem
Broadcasting in multihop packet radio network

Sharp Corp.

Location

Univ. of Osaka Prefecture
Univ. of Osaka Prefecture
Nippondenso Co. Ltd.

NTT Transmission System Lab
Kyoto Univ.

NTT Human Interface Lab
Univ. of Tokyo

Matsushita Electric

Yokohama National Univ.

Hitachi VLSI Engineering Corp. Kanazawa Univ.

Kobe Univ.

National Kisarazu Tech. College Univ. of Tokyo

Kobe Univ.

Hitachi Ltd.

Waseda Univ.

Kyoto Univ.

NTT Human Interface Lab Univ. of Tokushima

Hitachi Ltd.

Fukui Univ.
Kinki Univ.
Toshiba Corp.

Tamagawa Univ.

Tokyo Univ. of Agric. & Tech.

Tokyo Denki Univ.

Telecom Systems, Inc.

Sony Co., Ltd.

Univ. of Osaka Prefecture

Toyota Tech. Inst.

NTT Human Interface Lab
Keio Univ.

Table 5. Research Topics and Locations in Korea

Topic

Iterative autoassociative memory

ARMA model time series modeling
Parallel Boltzmann machine

Benchmarks for learning algorithms

3-D aircraft orientation identification

Image parameter estimation by error propagation

Hopfield network for self-tuning control

Functional approximation

Hopfield network for obstacle avoidance
Optical implementation for BAM

Perceptrons for image recognition
Hidden node reduction techniques
Distributed memory multiprocessors

Neural network for systolic design
Nearest neighbor classifier
Weight value initializations

Location

Kyungpook National Univ.

Korea Adv. Inst. of Sci. & Tech. Korea Adv. Inst. of Sci. & Tech. Pohang Inst. of Sci. & Tech. Kyungpook National Univ.

Korea Adv. Inst. of Sci. & Tech. Yonsei Univ.

Korea Adv. Inst. of Sci. & Tech. Kwangwoon Univ.

Korea Adv. Inst. of Sci. & Tech. Pohang Inst. of Sci. & Tech. Electronics & Telecomm Res. Inst. Korea Adv. Inst. of Sci. & Tech. Korea Adv. Inst. of Sci. & Tech. POSTECH

Korea Adv. Inst. of Sci. & Tech.

« iepriekšējāTurpināt »