Cooperation and modularity for classification through neural network techniques

We study modularity in the frame of neural network systems on two real-size applications. The cooperation of performant modules is used to improve recognition and rejection rates of handwritten digits coming from postal zip-codes. From a multi-class problem in multi-font character recognition, we have designed 49 neural submodules (one per class), and different cooperation schemes are studied and compared. The relation between the quality of the expert system and the efficiency of the cooperation scheme is shown.<
>