Invariant Feature Learning by Pairwise Neural Net Distance

Abstract

How to learn representations invariant to nuisance attribute is a universal problem among machine learning applications. This paper holds the following three contributions to this problem. First, we empirically show that adversarial training using categorical attribute classifier, which is one of the state-of-the-art approaches and called adversarial feature learning (AFL), is suffered from practical issues that significantly slow down the convergence. Second, we reformulate the optimization problem of AFL as pair-wise distribution matching and derive a new approach for learning invariant representations. Finally, we introduce parameter sharing techniques to reduce the computation difficulty of our strategy. Empirical results show the superior performance of our proposed method.