A stepwise procedure' for building and training a neural network intended to perform classification tasks, based on single layer learning rules, is presented. This procedure breaks up the classification task into subtasks of increasing complexity in order to make learning easier.The network structure is not fixed in advance: it is subject to a growth process during learning.Therefore, after training, the architecture of the network is guaranteed to be well adapted for the classification problem.

No abstract

It is shown that neural network classifiers with single-layer training can be applied efficiently to complex real-world classification problems such as the recognition of handwritten digits. The STEPNET procedure, which decomposes the problem into simpler subproblems which can be solved by linear separators, is introduced. Provided appropriate data representations and learning rules are used, performance comparable to that obtained by more complex networks can be achieved. Results from two different databases are presented: an European database comprising 8700 isolated digits and a zip code database from the US Postal Service comprising 9000 segmented digits. A hardware implementation of the classifier is briefly described.

We developed a check reading system, termed INTERCHEQUE, which recognizes both the legal (LAR) and the courtesy amount (CAR) on bank checks. The version presented here is designed for the recognition of French, omni-bank, omni-scriptor, handwritten bank checks, and meets industrial requirements, such as high processing speed, robustness, and extremely low error rates. We give an overview of our recognition system and discuss some of the pattern recognition techniques used. We also describe an installation which processes of the order of 70,000 checks per day. Results on a data base of about 170,000 checks show a recognition rate of about 75% for an error rate of the order of 1/10,000 checks.

We present a Neural Network -Hidden Markov ModelHybrid for the recognition of cursive words which are represented as left-right sequences of graphemes. The proposed approach models words with ergodic HMMs and is designed for small vocabularies. A single neural network provides grapheme observation probabilities for all HMMs in order to compute the most likely word model. During the iterative EM like training of the hybrid, the HMMs provide the targets for the discriminant training of the neural network. An extension of the approach to letter models which can be concatenated in order to form word models and which allow for large vocabularies is also briefly discussed. We report results obtained on a large data base of words from French cheques, showing recognition rates close to 93% for the 30 word vocabulary relevant for French legal amounts.