Multi-Layer Perceptron ensembles for increased performance and fault-tolerance in pattern recognition tasks

Multi-Layer Perceptrons (MLPs) have proven to be an e ective way to solve classi cation tasks. A major concern in their use is the di culty to de ne the proper network for a speci c application, due to the sensitivity to the initial conditions and to over tting and under tting problems which limit their generalization capability. Moreover, time and hardware constraints may seriously reduce the degrees of freedom in the search for a single optimal network. A very promising way to partially overcome such drawbacks is the use of MLP ensembles: averaging and voting techniques are largely used in classical statistical pattern recognition and can be fruitfully applied to MLP classi ers. This work summarizes our experience in this eld. A real-world OCR task is used as a test case to compare di erent models.