Bias: Bijective Input And Surjectivity In Zero Shot Learning

Zero-shot learning suffers from the issue of generalization due to domain shift across seen and unseen classes. In this paper, we propose a method that extends the usual approach of learning a mapping between semantic and visual embedding spaces by ensuring it to be surjective. This functional constraint along with triplet loss prevents the model from over-fitting to seen classes. We also use a bijective feature extractor to complement our proposal. Experimental results on benchmark datasets depict that our method outperforms standard approaches in conventional and generalized scenarios.