Distribution and gradient constrained embedding model for zero-shot learning with fewer seen samples

作者：

Highlights：

• A Distribution and Gradient constrained Embedding Model is proposed for zero-shot learning with fewer seen samples.

• We propose two regularizers to solve the representation bias problem and the over-fitting problem caused by limited samples.

• Extensive experiments show that DGEM outperforms the other baselines when each seen class only has 1/5 samples.

摘要

•A Distribution and Gradient constrained Embedding Model is proposed for zero-shot learning with fewer seen samples.•We propose two regularizers to solve the representation bias problem and the over-fitting problem caused by limited samples.•Extensive experiments show that DGEM outperforms the other baselines when each seen class only has 1/5 samples.

论文关键词：Zero-shot learning,Fewer seen samples,Representation bias,Over-fitting

论文评审过程：Received 20 January 2022, Revised 3 June 2022, Accepted 3 June 2022, Available online 22 June 2022, Version of Record 2 July 2022.

论文官网地址：https://doi.org/10.1016/j.knosys.2022.109218