Experiments with various recurrent neural network architectures for handwritten character recognition

This paper reports evaluations of several neural architectures when the handwritten character recognition is approached as a problem of spectro-temporal pattern recognition. In general, neural networks specialize in learning either the spectral or temporal characteristics of patterns. However, choice of appropriate features and architectures could lead to obtaining both spectral and temporal characteristics from the handwritten character patterns. One such feature and three appropriate architectures are the focus of this paper. The results obtained during a limited set of experiments indicate a great potential for the spectro-temporal approach to be a useful contender for being a part of schemes of handwritten character recognition systems. In addition, a simple voting method is presented for collaborative character recognition using three different recognition criteria.<>