A Local Learning Framework for Pattern Classification

The paper presents a local learning framework for pattern classification by partitioning a pattern space into different overlapped subsets and combining decisions in a local space. In contrast to designing a classifier on the global space, the advantage of the local learning framework is to reduce the complexity of component classifier which helps to enhance generalization. Our experimental results on handwritten digit database (CENPARMI and MNIST) are comparable with current best classifiers, which indicates that the proposed method is effective for classifying real world patterns. In addition, we have analyzed the characteristics of overlapped subsets and discovered that a simple average of the outputs of component classifiers achieves the best performance by optimizing the weights. Keywords— Local learning, classifier complexity, pattern classification, generalization