Trusted Deep Domain Adaptation with Uncertainty Measure Based on Evidence Theory

Abstract

Domain adaptation aims to build up an adaptive model for the learning tasks on target domain by using the data from source domain. Existing domain adaptation methods focused on reducing the gap between the data distributions of source domain and target domain but neglected the uncertainty of source domain data for the cross-domain learning. Actually, there exit the data in source domain that are uncertain with respect to the learning task on target domain. The uncertainty will cause negative transfer impacts and lead to the untrustworthy domain adaptation. Aiming at the problem, we propose a measure based on evidence theory to represent the uncertainty of source domain data for cross-domain classification and utilize the uncertainty measure to construct a trusted deep adaptation neural network (TDAN) to implement the trustworthy domain adaptation. Specifically, we design an evidential mass function to represent the uncertainty of source domain data with respect to the classification task on target domain, and formulate the loss function of the trusted deep adaptation network with Dirichlet distribution to involve the uncertainty. Experiments on real-world cross-domain learning tasks validate the trusted deep adaptation network can handle the uncertain data in source domain and produce trustworthy learning results of domain adaptation.