Local Logistic Classifiers for Large Scale Learning

This article describes the construction of the local logistic classifier, a highly scalable and parallelizable classifier combining local learning with a logistic classifier. To speed up the learning of each local logistic classifier, a novel fixed point algorithm is introduced. The practicality and performance of the classifier is demonstrated on a 10 million character training set derived from the UW3 dataset, and the classifier is demonstrated as part of an OCR system. The classifier is also compared to another proposed large scale classification method, GURLS [7], in experiments on ImageNet dataset .