SEER-ZSL: Semantic Encoder-Enhanced Representations for Generalized Zero-Shot Learning

Abstract

Zero-Shot Learning (ZSL) presents the challenge of identifying categories not seen during training. This task is crucial in domains where it is costly, prohibited, or simply not feasible to collect training data. ZSL depends on a mapping between the visual space and available semantic information. Prior works learn a mapping between spaces that can be exploited during inference. We contend, however, that the disparity between meticulously curated semantic spaces and the inherently noisy nature of real-world data remains a substantial and unresolved challenge. In this paper, we address this by introducing a Semantic Encoder-Enhanced Representations for Zero-Shot Learning (SEER-ZSL). We propose a hybrid strategy to address the generalization gap. First, we aim to distill meaningful semantic information using a probabilistic encoder, enhancing the semantic consistency and robustness. Second, we distill the visual space by exploiting the learned data distribution through an adversarially trained generator. Finally, we align the distilled information, enabling a mapping of unseen categories onto the true data manifold. We demonstrate empirically that this approach yields a model that outperforms the state-of-the-art benchmarks in terms of both generalization and benchmarks across diverse settings with small, medium, and large datasets. The complete code is available on GitHub.