Introduction

Abstract

The objective of this book is to establish a framework which combines two different paradigms in machine learning: global learning and local learning. The combined model demonstrates that a hybrid learning of these two different schools of approaches can outperform each isolated approach both theoretically and empirically. Global learning focuses on describing a phenomenon or modeling data in a global way. For example, a distribution over the variables is usually estimated for summarizing the data. Its output can usually reconstruct the data. This school of approaches, including Bayesian Networks [8, 13, 30], Gaussian Mixture Models [3, 21], and Hidden Markov Models [2, 25], has a long an distinguished history, which has been extensively applied in artificial intelligence [26], pattern recognition [9], and computer vision [7]. On the other hand, local learning does not intend to summarize a phenomenon, but builds learning systems by concentrating on some local parts of data. It lacks the flexibility yet surprisingly demonstrates superior performance to global learning according to recent researches [4, 16, 15]. In this book, a bridge has been established between these two different paradigms. Moreover, the resulting principled framework subsumes several important models, which respectively locate themselves into the global learning paradigm and the local learning paradigm.