Discriminative Writer Adaptation

10th Intl. Workshop on Frontiers in Handwriting Recognition (IWFHR) |

We propose a general method for adapting a writer-independent classifier to an individual writer. We employ a mixture of experts formulation, where the classifiers are trained on weighted clusters of writers. The clusters are determined by which experts classify individual writing correctly. The method adapts by choosing the appropriate combination of classifiers for a new user. It applies to any probabilistic discriminative classifier, and adapts discriminatively without modeling the input feature distribution. We apply the method to online character recognition. Specifically, we use a mixture of neural networks as well as a mixture of logistic regressions. We train the mixture via conjugate gradient ascent or via the EM algorithm on 192,000 Latin characters of 98 classes and 216 writers, and show adaptation results for 21 writers.