|
Abstract : |
We have developeda neural-network architecture for recognizing handwritten digits. This network has 1 % error rate with about 7% reject rate on handwritten zipcode digits provided by the U.S. Postal Service. In this paper, we discussimplementing this architecture in a real-world character recogmtion system. The main issue is the trade-off between cost and benefits such as accuracy and speed. A method for combiningindependentlytrained networks to achievehigher per-formance at relatively low cost is presented. Accurate estimatesof the probability of correct recognition, aswell asrunner-upprobabilities, areof ever-increasingimportanceasrecog-nition systems move out of the lab into the real world. Per-character probabilitiesgiveus the information necessaryfor calculatingper-field or multi-field probabilities. Wediscussa methodfor normalizingout-put activations levels,thus providing a normalizedscore,which (for our network at least) is a good estimate of the probability. We also find that using this normalizedscoreasa rejection thresholdgivessimilar performance to previous rejection schemes. We also discussthe important and complex relationship between rejection threshold, averagenumber of errors, and the cost of errors., |
