Learning to Separate Domains in Generalized Zero-Shot and Open Set Learning: a probabilistic perspective

Hanze Dong, Yanwei Fu, Leonid Sigal, SungJu Hwang, Yu-Gang Jiang, Xiangyang Xue

Abstract: This paper studies the problem of domain division which aims to segment instances drawn from different probabilistic distributions. This problem exists in many previous recognition tasks, such as Open Set Learning (OSL) and Generalized Zero-Shot Learning (G-ZSL), where the testing instances come from either seen or unseen/novel classes with different probabilistic distributions. Previous works only calibrate the confident prediction of classifiers of seen classes (WSVM Scheirer et al. (2014)) or taking unseen classes as outliers Socher et al. (2013). In contrast, this paper proposes a probabilistic way of directly estimating and fine-tuning the decision boundary between seen and unseen classes. In particular, we propose a domain division algorithm to split the testing instances into known, unknown and uncertain domains, and then conduct recognition tasks in each domain. Two statistical tools, namely, bootstrapping and KolmogorovSmirnov (K-S) Test, for the first time, are introduced to uncover and fine-tune the decision boundary of each domain. Critically, the uncertain domain is newly introduced in our framework to adopt those instances whose domain labels cannot be predicted confidently. Extensive experiments demonstrate that our approach achieved the state-of-the-art performance on OSL and G-ZSL benchmarks.

Keywords: Generalized zero-shot learning, domain division, bootstrapping, Kolmogorov-Smirnov

TL;DR: This paper studies the problem of domain division by segmenting instances drawn from different probabilistic distributions.

Data: [AwA](https://paperswithcode.com/dataset/awa-1), [aPY](https://paperswithcode.com/dataset/apy)