摘要
与传统零样本识别相比,广义零样本识别的样本不仅包括测试类别样本,还包括训练类别样本,因此,广义零样本识别更具有现实意义。提出了一种基于混合高斯分布的广义零样本识别的算法(MGM VAE),在编码器中采用多个通道结构,促使变分自编码器(VAE)模型可以在更广泛的空间内寻求更好的映射解。
Compared with the traditional zero-shot learning,generalized zero-shot learning includes the test category and the training category.Therefore,generalized zero-shot learning is more realistic.This paper proposes a generalized zero-shot learning algorithm based on a Gaussian mixture distribution(MGM-VAE).Multi-channel structures is used in the encoder,so that the variational auto encoding(VAE)model can seek a better mapping solution in a wider space.
作者
邵洁
李晓瑞
SHAO Jie;LI Xiaorui(School of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《上海电力大学学报》
CAS
2021年第5期475-480,共6页
Journal of Shanghai University of Electric Power
关键词
广义零样本识别
混合高斯模型
变分自编码器
generalized zero-shot learning
Gaussian mixture
variational auto encoding
