Gram-Schmidt orthogonalization neural nets for OCR

A description is given of a three-layer neural network for pattern classification/character recognition. The first layer is a heteroassociative feedforward network with bipolar output (+or-1) and zero threshold neurons. The second layer is an autoassociative memory whose input-output characteristics are the same as those in the first layer. The third layer is used to recognize the pattern and control whether the new orthogonal feature vector should be installed by the outer product formula to increase the memory capacity to M´=M+1. With this network, conventional pattern recognition of the minimax type is used to determine the initial interconnection matrix. The samples are classified by means of supervised learning. Only a single physical layer need be built, since the same layer can repeatedly be used three times in series. The performance of the network is studied.<>