Cross-Layer Autoencoder for Zero-Shot Learning

Abstract

Zero-shot learning (ZSL) is the task of recognizing samples from their related classes which have never been seen during model training. ZSL is generally realized through learning a common embedding space where both high dimensional visual features and some pre-defined semantics can be mapped. However, this kind of solutions usually suffers from domain shift. In addition, the limitation and subjectivity of manual semantic information can also affect the classification results. To address these challenges, this paper proposes a novel end-to-end deep learning model called Cross-Layer Autoencoder (CLAE), which integrates different ways of semantic mapping and maintains reconstruction information. Besides, a regularized loss function is used to preserve local class manifolds. Extensive experiments for both traditional and generalized ZSL tasks are conducted on several benchmark datasets, and high effectiveness of the proposed method and its superiority over many previous researches are demonstrated. INDEX TERMS Zero-shot learning, cross-layer autoencoder, regularized loss function, class manifolds.

Zero-Shot Image Classification Based on a Learnable Deep Metric

“…The space embedding methods [ 8 , 9 , 10 ] rely on an embedding space, in which the attribute migration from seen classes to unseen classes is completed. The generative model methods utilize different generative models, such as generative adversarial networks (GANs) [ 16 , 17 ], variational autoencoders (VAEs) [ 18 , 19 ], and flow-based models (Flows) [ 20 ] to directly generate visual features of unseen classes, and then transform the zero-shot learning problem into a traditional supervised learning problem. For example, [ 16 ] has proposed a triple discriminator GAN (TDGAN), which employs a GAN with three discriminators to synthesize visual features for images of unseen classes.…”