Output details

A theoretical framework for multiple neural network systems

<22>Multiple classifier systems perform on average better than single classifiers, as evidenced through ensembles of neural networks. However, despite empirical results demonstrating performance gains from ensembles, there is no theoretical work which says which combinations of classifiers are best. This paper for the first time applies rigorous mathematical concepts to the properties of multiple neural networks. The results establish that a suitable choice of functions can allow properties of the whole system to be inferred irrespective of whether the system is a single neural network or an ensemble. This is the first theoretical result linking these types of system.