MLP modular networks for multi-class recognition.
Philippe Sebire, Bernadette Dorizzi · 1993
We present a connectionist modular approach which is potentially able to deal with real-size applications as its size does not increase drastically with the size of the problem. It relies on very simple cooperation schemes of modular MLP networks especially designed for some sub-tasks. Several cutting up are tested : from two or three nets to one network per class. These approaches are compared on a multi-class classification task (recognition of typographic characters) in terms of performance rates. 1. Introduction Multi Layer Perceptron (MLP) networks have been extensively studied during the last years. In many applications (speech or handwritten character recognition, time series prediction etc...) they have proved to be very efficient classifiers, especially because they do not necessitate a lot of preprocessings. Most often, these studies have been conducted on small size problems with a limited number of classes, which presents the advantage of maintaining the simulation time rea...