Domain Adaptation: Learning Bounds and Algorithms - Computer Science > LearningReportar como inadecuado

Domain Adaptation: Learning Bounds and Algorithms - Computer Science > Learning - Descarga este documento en PDF. Documentación en PDF para descargar gratis. Disponible también para leer online.

Abstract: This paper addresses the general problem of domain adaptation which arises ina variety of applications where the distribution of the labeled sampleavailable somewhat differs from that of the test data. Building on previouswork by Ben-David et al. 2007, we introduce a novel distance betweendistributions, discrepancy distance, that is tailored to adaptation problemswith arbitrary loss functions. We give Rademacher complexity bounds forestimating the discrepancy distance from finite samples for different lossfunctions. Using this distance, we derive novel generalization bounds fordomain adaptation for a wide family of loss functions. We also present a seriesof novel adaptation bounds for large classes of regularization-basedalgorithms, including support vector machines and kernel ridge regression basedon the empirical discrepancy. This motivates our analysis of the problem ofminimizing the empirical discrepancy for various loss functions for which wealso give novel algorithms. We report the results of preliminary experimentsthat demonstrate the benefits of our discrepancy minimization algorithms fordomain adaptation.

Autor: Yishay Mansour, Mehryar Mohri, Afshin Rostamizadeh


Documentos relacionados