Generalized Zero-Shot Learning using Identifiable Variational Autoencoders

Deep learning tasks rely heavily on a large amount of training data, but collecting and annotating data daily is not practical. Therefore, Zero-shot learning (ZSL) has become important for the applications, where no labeled data is available during training. ZSL aims at recognizing unseen classes by semantic transfer of information from seen to unseen classes. In this paper, we have proposed an identifiable VAE (iVAE) based generative model to address conventional and generalized ZSL. The key to our approach is learning disentangled representations, where each dimension is statistically independent and responsible for generating data. Thus, VAE is a commonly used model for learning disentangled independent factors of variation from the data. Our goal is to learn a latent space representing significant information, that approximates the actual data distribution. Extensive experiments on five benchmark datasets, i.e. CUB, AWA1, AWA2, SUN and aPY, are performed for further evaluation in both settings.