Improved method of handwritten digit recognition tested on MNIST database

We have developed a novel neural classifier LImited Receptive Area (LIRA) for the image recognition. The classifier LIRA contains three neuron layers: sensor, associative and output layers. The sensor layer is connected with the associative layer with no modifiable random connections and the associative layer is connected with the output layer with trainable connections. The training process converges sufficiently fast. This classifier does not use floating point and multiplication operations. The classifier was tested on two image databases. The first database is the MNIST database. It contains 60,000 handwritten digit images for the classifier training and 10,000 handwritten digit images for the classifier testing. The second database contains 441 images of the assembly microdevice. The problem under investigation is to recognize the position of the pin relatively to the hole. A random procedure was used for partition of the database to training and testing subsets. There are many results for the MNIST database in the literature. In the best cases, the error rates are 0.7, 0.63 and 0.42%. The classifier LIRA gives error rate of 0.61% as a mean value of three trials. In task of the pin-hole position estimation the classifier LIRA also shows sufficiently good results. (C) 2004 Elsevier B.V. All rights reserved.