An application of neural net chips: handwritten digit recognition

A general-purpose, fully interconnected neural-net chip was used to perform computationally intensive tasks for handwritten digit recognition. The chip has nearly 3000 programmable connections, which can be set for template matching. The templates can be reprogrammed as needed during the recognition sequence. The recognition process proceeds in four major steps. First, the image is captured using a TV camera and a digital framegrab. This image is converted, using a digital computer, to either black or white pixels and scaled to fill a 16*16-pixel frame. Next, using the neural-net chip, the image is skeletonized, i.e. the image is thinned to a backbone one pixel wide. Then, the chip is programmed, and a feature map is created by template-matching stored primitive patterns on the chip with regions on the skeletonized image. Finally, recognition, based on the feature map, is achieved using any one of a variety of statistical and heuristic techniques on a digital computer. Best scores range between 90% and 99% correct classification, depending on the quality of the original handwritten digits.<>