Towards Adversarial Robustness and Reducing Uncertainty Bias through Expert Regularized Pseudo-Bidirectional Alignment in Transductive Zero Shot Learning

Abhishek Kumar Sinha, Deepak Mishra, S. Manthira Moorthi

Abstract: Transductive zero-shot learning (TZSL) aims to minimize the domain shift between the learned and true distribution of the unseen classes by allowing access to the unpaired samples from unseen classes. While many distribution alignment based methods attempt to align both visual and semantic spaces to train the classifier, their performance is still limited by confirmation bias. Additionally, bidirectional alignment approaches are based on the strong assumption that the intrinsic dimensions of visual and semantic spaces are the same, which is rarely true. In this work, we first highlight the limitations of bidirectional alignment in terms of intrinsic dimensionality. We then present a pseudo-bidirectional approach that, without any underlying assumptions on these spaces, utilizes the learned visual-to-attribute mapping to minimize the distribution shift between learned and true unseen visual feature distributions. We further utilize an entangled loss between semantic and visual space to minimize the confirmation or uncertainty bias and improve the adversarial robustness. We, theoretically and empirically, show the performance gain in addition to the adversarial robustness under the proposed setting.