Title :
A fast algorithm for the hidden units in a multilayer perceptron
Author_Institution :
Graphic Syst. Res. Lab., Matsushita Electr. Ind. Co. Ltd., Tokyo, Japan
Abstract :
We propose an online algorithm for deciding the smallest possible number of neurons in a hidden layer of the backpropagation model for the best possible operation based on singular value decomposition. Our simulations showed that the algorithm proposed works very well for character recognition. We conclude that the algorithm can be a powerful tool for solving a general classification problem with an optimized neural network.
Keywords :
backpropagation; character recognition; multilayer perceptrons; optimisation; singular value decomposition; backpropagation model; character recognition; fast algorithm; hidden layers; multilayer perceptron; neural network; neurons; singular value decomposition; Algorithm design and analysis; Character recognition; Convergence; Covariance matrix; Handwriting recognition; Multi-layer neural network; Multilayer perceptrons; Neural networks; Neurons; Pattern recognition;
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.1993.713926
