Monday, November 6
2:00pm,  STII Room 330B

Some Recent Advances in Near-neighbor Statistical Learning

Abstract

Classification is the problem of labeling an unknown test sample given a database of learned samples. For example, one might classify a patient as having diabetes, or classify an utterance as the digit "one". Arguably the most intuitive approach to classification is to find a sample in the database that is most similar to the test sample, and give it the same label. This approach is known as nearest-neighbor classification. In this talk, we discuss some recent advances in nearest-neighbor learning, in particular how to weight a set of nearest-neighbors to form a weighted vote about the class label, and how to choose a set of nearest-neighbors from the database. Specifically, it is shown that weights that solve linear interpolation equations minimize the first-order learning error, and when coupled with the principle of maximum entropy this results in significantly improved nearest-neighbor classification performance. New approaches to adaptively determining neighborhoods for local learning that enclose the test point are also discussed. A number of applications are discussed briefly, including finding corrosion in pipelines, classifying audio signals, and color management.

Speaker Bio