Input driven MLP model and its application to the recognition of Chinese characters
Henry Shu-Hung Chung, Jin Kyung Ryeu, W.I. Lee · 1991
Proposes a new neural network model, called the input driven multilayer perceptron (IDMLP), and its learning algorithm for binary input and output patterns. In contrast to the back propagation algorithm (BPA), hidden layers develop by themselves as the learning algorithm proceeds. The learning speed is much faster than the BPA. Hard limiters as the activation functions of neurons and integer connection weights are used. Accurate hardware implementation of trained networks can be easily realized using readily available digital CMOS VLSI technology. As practical applications IDMLPs are trained to recognize sets of randomly chosen printed Chinese characters after a feature extraction process. Discussions on actual digital CMOS VLSI implementation of trained IDMLPs are are also included.>