AbstractNeural network technology, especially that related to or derived from the multilayered Perceptron model with the back‐propagation learning mechanism, has been found to be effective in pattern recognition and classification. Although mentioned infrequently, the number of patterns that can be recognized or classified by a specific neural network based on this technology is inescapably constrained by the network configuration such as the numbers of input, output, and hidden units. A direct expansion, e.g., increasing the numbers of the various units, of an existing network to accommodate additional patterns is usually undesirable for two reasons. First, a complete retraining of the network, incorporating the learned and new patterns as the training set, is necessary. Second, such a retraining almost always proceeds at a decreasing rate and is more difficult to converge. Both of these problems can be circumvented, at least partially, by resorting to the concept of modular neural networks. In the present work, each of two multilayered neural networks is trained to recognize eight types of starches used in foods. Subsequently, the two trained networks are coupled as modules of a final network. The learning rate of the final network is found to be much faster than that of a nonmodularized network with the same configuratio
展开▼
