Subject: ICR Presents a Colloquium on "Hybrid-Genetic/Gradient Learning in Multi-Layer Neural Networks", Wednesday, November 25, 1992 at 3:30 pm in DC 1302.
Message-ID: <BxvBzs.Az@watserv1.uwaterloo.ca>
Keywords: Speaker: Dr. Mohamad H. Hassoun, Computation and Neural Network Lab., Dept. of Elec. and Comp. Eng., Wayne State Univ., Detroit, Michigan.
Sender: news@watserv1.uwaterloo.ca
Organization: University of Waterloo
Distribution: ont
Date: Tue, 17 Nov 1992 16:19:03 GMT
Lines: 80
The University of Waterloo
200 University Avenue
Waterloo, Ontario
The Institute for Computer Research (ICR)
Presents a Colloquium on
Hybrid Genetic/Gradient Learning in Multi-Layer Neural Networks
by: Dr. Mohamad H. Hassoun
of: Computation and Neural Network Laboratory
Department of Electrical and Computer Engineering
Wayne State University
Detroit, Michigan
Date: Wednesday, November 25, 1992
Time: 3:30 p.m.
Place: William G. Davis Computer Research Centre, Room 1302
Abstract
Learning in neural networks is a form of search in a multi-
dimensional space. The popular backpropagation learning algo-
rithm for layered neural networks represents a class of local
gradient-based search. The optimality of gradient search in com-
plex spaces (e.g., neural net weight spaces) is constrained due
to local minimas. On the other hand, global search strategies
(simulated annealing, genetic algorithms, etc.) lead to optimal
solutions but at the expense of increased computational
complexity/search time. In this presentation, a novel learning
technique is proposed for multi-layer neural networks based on a
hybrid genetic search in hidden target space/gradient search in
weigh space strategy. Several versions of the proposed algorithm
are evaluated on benchmark problems with varying complexity. Our
simulations show that the new algorithm combines the global op-
timization capabilities of genetic algorithms with the speed of
gradient descent search, for increased learning efficiency. In
addition, we show that genetic search in hidden target space is
less complex than that in weight space.
Biography:
Mohamad H. Hassoun was born in Lebanon, in 1961. He received the
B.S., M.S., and Ph.D. degrees in electrical engineering from
Wayne State University (WSU), Detroit, Michigan, in 1981, 1982,
and 1986, respectively. He is currently as Associate Professor
in the Department of Electrical and Computer Engineering at Wayne
State University. Dr. Hassoun's research interests are in artif-
icial neural systems, parallel collective optimization and compu-
tations, learning algorithms, and optical and electronic neural
network implementations.
Dr. Hassoun has received a National Science Foundation Research
Initiation Award and a Presidential Young Investigator Award in
1988 and 1990, respectively. He is presently the associate edi-
tor for book reviews and a technical associate editor of the IEEE
Transactions on Neural Networks. He currently serves as program
committee member for the 1993 IEEE International Conference on
Neural Networks - ICNN (San Francisco, CA, March 1993) and the
International Workshop on Artificial Neural Networks - IWANN
(Spain, June 1993). He has published over 35 journal and confer-
ence proceedings papers on the subject of neural networks. Dr.
Hassoun is a member of IEEE, INNS, SPIE, Sigma Xi, ASEE, and Tau
