Over the last several years, computing systems based on adaptive learning with fine-grained parallel architectures have moved from obscurity to front-page prominence. These systems derive some of their novel architecture from ideas gleaned from biology, hence the name “neural network”. Although many of the ideas behind this field are not new, improved computing hardware, better understanding of learning algorithms, and limitations of traditional approaches have combined to renew interest in neural nets.

The ultimate success of electronic neural networks will depend on their effectiveness in solving real-world problems. Therefore it is important to choose realistic benchmarks as a focus for research in algorithms and hardware for neural-network computing. Optical character recognition (OCR) of handwritten digits is such a benchmark problem: it has a clearly defined commercial importance and a level of difficulty that makes it challenging, yet it is not so large as to be completely intractable.

We have demonstrated that a neural net can perform handwritten digit recognition with state-of-the-art accuracy. The solution required “automatic learning” and generalization from thousands of training examples, and also required designing into the system considerable knowledge about the task — neither engineering alone nor learning from examples alone would have sufficed. The resulting network is well-suited for implementation on workstations or PCs, and can take advantage of digital signal processors (DSPs) or custom VLSI.

Download pdf Optical Character Recognition A Technology Driver for Neural Networks