Optimal Bayesian Transfer Learning

doi:10.1117/3.2540669.ch6

Book: Optimal Bayesian Classification

Author(s): Lori A. Dalton, Edward R. Dougherty

Published: 2020

https://doi.org/10.1117/3.2540669.ch6

Author Affiliations +

Abstract

The theory of optimal Bayesian classification assumes that the sample data come from the unknown true feature-label distribution, which is a standard assumption in classification theory. When data from the true feature-label distribution are limited, it is possible to use data from a related feature-label distribution. This is the basic idea behind transfer learning, where data from a source domain are used to augment data from a target domain, which may follow a different feature-label distribution (Pan and Yang, 2010; Weiss et al., 2016). The key issue is to quantify relatedness, which means providing a rigorous mathematical framework to characterize transferability. This can be achieved by extending the OBC framework so that transfer learning from the source to target domain is via a joint prior distribution for the model parameters of the feature-label distributions of the two domains (Karbalayghareh et al., 2018b). In this way, the posterior distribution of the target model parameters can be updated via the joint prior probability distribution function in conjunction with the source and target data.