Prototype adjustment for zero shot classification

Abstract

Zero shot classification addresses the problem of classifying unseen classes with seen class samples. Current zero shot learning methods mostly focus on learning the mapping function from image feature space to semantic space which is extremely important. However, these methods assume the seen and unseen class prototypes are fixed. A class prototype is referred to the semantic representation of a class. The semantic representation is represented by the attributes or word vectors which may be inaccurate and not discriminative. We attempt to find new prototypes that are more accurate for the zero shot classification tasks. In this paper, we proposed a Prototype adjustment method for the zero shot classification tasks (PAZSC) by adjusting the prototypes and learning the mapping function from image feature space to semantic space, simultaneously. The adjusted prototypes are more separable and discriminative for the zero shot classification tasks. A joint optimization function is proposed to learn the new prototypes and the mapping function. What is more, there is a domain shift problem in zero shot classification tasks caused by the disjointed seen and unseen images. We further learn a more generalizable mapping function to alleviate the domain shift problem. We have experimented on the state-of-the-art zero shot learning datasets, demonstrating that our PAZSL method has good performance.