Unseen class detection via Ensemble uncertainty for Generalized Zero-shot Learning
Unseen class detection via Ensemble uncertainty for Generalized Zero-shot Learning

概要

Generalized zero-shot learning (GZSL) is a challenging topic in classification tasks. GZSL, provided auxiliary information about class labels, aims to classify not only samples from seen
classes but also samples from classes unlearned during training. However, in GZSL, a learned classifier would encounter a problem that the outputs of the classifier intend to bias toward seen
classes when applying the conventional zero-shot learning (ZSL) method to the GZSL framework. To improve this problem, we propose an ensemble method that estimates uncertainty for a test
sample and infers whether a test sample is from a seen class or an unseen class according to the estimated uncertainty, so we can select the proper search space of class labels for the test sample on the basis of the inference. In doing so, the samples judged as from seen classes can be classified as a normal classification task and the samples judged as from unseen classes can be applied to any existing ZSL method without encountering the bias problem. We conducted experiments on benchmark datasets including AWA and CUB and showed that classification performance could be improved ved significantly

産業界への展開例・適用分野

pattern recognition, machine learning, zero-shot learning

研究者

氏名	専攻	研究室	役職/学年
Chen zhaozhi	システム科学専攻	情報数理システム	博士1回生