Handwritten Digit Recognition: Applications of Neural Net Chips and Automatic Learning

Abstract

We describe two neural-net approaches to digit recognition. One method uses a neural-network chip to perform line thinning and local feature extraction. This preprocessing stage was designed by hand and did not involve any learning. However, automatic learning was used in the final classification step. The chip can process about 100 characters/sec, but the interface to the host computer limits the throughput to about 1 character/sec.

The other method uses constrained automatic learning on pixel images with no preprocessing other than segmentation and size-normalization. It appears that good generalization performance cannot be obtained unless some a priori knowledge about the task is built into the system. This paper demonstrates how such knowledge can be integrated into a back-propagation network by providing hints and constraints on the architecture. Most of the computational burden of this method can be borne by a digital signal processor, resulting in a throughput of about 10 characters/sec.

Both methods have 1% error rate and about a 12–13% reject rate when trained on 7300 digits taken from U.S. Zipcodes and tested on 2000 such digits. These results appear to be at the state-of-the-art in handwritten digit recognition.