An Integral Projection-based Semantic Autoencoder for Zero-Shot Learning

Zero-shot Learning (ZSL) classification categorizes or predicts classes (labels) that are not included in the training set (unseen classes). Recent works proposed different semantic autoencoder (SAE) models where encoder embeds a visual feature vector space into and decoder reconstructs original space. The objective is to learn embedding by leveraging source data distribution, which can be applied effectively but related target distribution. Such embedding-based methods prone domain shift problems vulnerable biases. We propose an integral projection-based (IP-SAE) projects concatenated with latent representation force reconstruct visual-semantic Due this constraint, projection function preserves discriminatory inside enriched forces more precise reconstitution of invariant manifold. Consequently, learned less domain-specific alleviates problem. Our IP-SAE model consolidates symmetric transformation for projection, thus, it provides transparency interpreting generative applications ZSL. Therefore, addition outperforming state-of-the-art considering four benchmark datasets, our analytical approach allows us investigate distinct characteristics generative-based unique context zero-shot inference.