Machine Learning in Non-Stationary Environments: Introduction to Covariate Shift Adaptation

摘要: As the power of computing has grown over past few decades, field machine learning advanced rapidly in both theory and practice. Machine methods are usually based on assumption that data generation mechanism does not change time. Yet real-world applications learning, including image recognition, natural language processing, speech robot control, bioinformatics, often violate this common assumption. Dealing with non-stationarity is one modern learning's greatest challenges. This book focuses a specific non-stationary environment known as covariate shift, which distributions inputs (queries) but conditional distribution outputs (answers) unchanged, presents theory, algorithms, to overcome variety non-stationarity. After reviewing state-of-the-art research field, authors discuss topics include under model selection, importance estimation, active learning. They describe such real world shift adaption brain-computer interface, speaker identification, age prediction from facial images. With book, they aim encourage future statistics, engineering strives create truly autonomous machines able learn