Pseudo distribution on unseen classes for generalized zero shot learning.

•Attribute similarity is exploited as the Pseudo distribution to solve the over-fitting on GZSL;•Attribute similarity is further compressed as one-hot vector to encourage the certainty of the training;•Visual space is employed as the embedding space to alleviate the hubness problem;•The proposed PSD can significantly outperform the SOTA methods by large margins on GZSL.