Multiresolution Learning on Neural Network Classifiers: A Systematic Approach

Abstract

One of the most crucial challenges for classifiers is generalization. In this paper, we present a novel and systematic multiresolution learning approach for neural network classifiers to improve their generalization performance on classification tasks with feature based input space. The proposed approach adopts agglomerative hierarchical clustering to generate coarser resolution training data from the original data because hierarchical clustering captures the detailed structure of clustering for any given data set in feature space, and an effective algorithm is developed to automatically extract critical coarser resolution levels for each class. The proposed approach is thoroughly evaluated through experiments on six real-world benchmark data sets, where traditional learning (i.e., single resolution learning) is used as the baseline. The empirical results demonstrate that multiresolution learning significantly improves neural network classifiers' generalization performance when compared to the baseline, especially for very difficult tasks.