Title :
An empirical paralogism of the backpropagation networks
Author_Institution :
Dept. of Manage. Inf. Syst., Nat. Chengchi Univ., Taipei, Taiwan
Abstract :
In the learning process of the backpropagation networks, the numerical phenomenon of being likely to have a sluggish processing somewhere and to terminate there is usually interpreted as being trapped by a relatively minimal point. However, this claim might be wrong because this numerical phenomenon might also happen in the vicinity of a saddle stationary point.
Keywords :
backpropagation; feedforward neural nets; numerical analysis; optimisation; attractor; backpropagation networks; empirical paralogism; feedforward neural network; learning process; minimal point; numerical phenomenon; saddle stationary point; Eigenvalues and eigenfunctions; Electronic mail; Entropy; Feedforward systems; Gradient methods; Iterative algorithms; Management information systems; Optimization methods;
Conference_Titel :
Neural Networks, 1993. IJCNN '93-Nagoya. Proceedings of 1993 International Joint Conference on
Print_ISBN :
0-7803-1421-2
DOI :
10.1109/IJCNN.1993.714209
