Robust Zero-Shot Learning with Distribution-Preserving Feature Generation and Bias-Calibrated Classification

Abstract

The ability to recognize objects out of our environment with the use of additional information such as attribute vectors or text embeddings-this is what people call Zero-Shot Learning. Believe me, there are many more ZSL techniques that get wrong in real-life context - they're not always semantically correct with visuals, can bias towards existing classes of visuals, and don't model features that well. That takes a toll on the task of generalizing. To address those problems, this paper proposes a hybrid ZSL model based on the combination of transformer-based semantic embedding, generative feature synthesis based on flow and bias-calibrated classification. The answer is, here: The transformer module aims to stitch together the relationships between semantic attributes and visual features. The flow based generative model has the capability to synthesize features for out of vocabulary classes that retains both the form and distribution of features, giving stable and life-like features. Plus, the bias calibration method yields a fair prediction for unseen classes, which is important in a generalized zero-shot settings. It is a multi-stage optimization framework that integrates the representation learning and feature generation. Results on large datasets such as CUB, AwA2 and SUN demonstrate clear improvements on the unseen classes both in terms of the harmonic mean as well as accuracy against state-of-the-art approaches. In a word, this method is sound, robust, and scalable - it is up to tackle real-world ZSL problems.