Design and implementation of a dedicated neural network for handwritten digit recognition

The automatic recognition of handwritten digits seems to be one of the most promising fields for applications of artificial neural networks; various studies have shown that good recognition rates can be obtained on large ´real-world´ data bases. This paper presents: (i) the design of a network architecture, resulting from a stepwise procedure developed at ESPCI for simultaneously building and training a neural network, intended for the automatic recognition of isolated handwritten digits and (ii) a silicon implementation of that network, using a general-purpose neural circuit architecture developed at CSI.<>