Exploring the Limits of Zero-Shot Learning - How Low Can You Go?
Standard zero-shot learning (ZSL) methods use a large number of seen categories to predict very few unseen categories while maintaining unified data splits and evaluation metrics. This has enabled the research community to advance notably towards formulating a standard benchmark ZSL algorithm. However, the most substantial impact of ZSL lies in enabling the prediction of a large number of unseen categories from very few seen categories within a specific domain. This permits the collection and annotation of training data for only a few previously seen categories, thereby significantly mitigating the training data collection and annotation process. We address the difficult problem of predicting a large number of unseen object categories from very few previously seen categories and propose a framework that enables us to examine the limits of inferring several unseen object categories from very few previously seen object categories, i.e., the limits of ZSL. We examine the functional dependence of the classification accuracy of unseen object classes on the number and types of previously seen classes and determine the minimum number and types of previously seen classes required to achieve a prespecified classification accuracy for the unseen classes on three standard ZSL data sets. An experimental comparison of the proposed framework to a prominent ZSL technique on these data sets shows that the proposed framework achieves higher classification accuracy on average while providing valuable insights into the unseen class inference process.