In an optical character recognition problem, we compare (as a function of training set size) the performance of three neural network based ensemble methods (two versions of boosting and a committee of neural networks trained independently) to that of a single network. In boosting, the number of patterns actually used for training is a subset of all potential training patterns. Based on either a fixed computational cost or training set size criterion, some version of boosting is best. We also compare (for a fixed training set size) boosting to the following algorithms: optimal margin classifiers, tangent distance, local learning, k-nearest neighbor, and a large weight sharing network with the boosting algorithm showing the best performance.
展开▼