Most existing zero-shot learning methods consider the problem as a visualsemantic embedding one. Given the demonstrated capability of GenerativeAdversarial Networks(GANs) to generate images, we instead leverage GANs toimagine unseen categories from text descriptions and hence recognize novelclasses with no examples being seen. Specifically, we propose a simple yeteffective generative model that takes as input noisy text descriptions about anunseen class (e.g.Wikipedia articles) and generates synthesized visual featuresfor this class. With added pseudo data, zero-shot learning is naturallyconverted to a traditional classification problem. Additionally, to preservethe inter-class discrimination of the generated features, a visual pivotregularization is proposed as an explicit supervision. Unlike previous methodsusing complex engineered regularizers, our approach can suppress the noise wellwithout additional regularization. Empirically, we show that our methodconsistently outperforms the state of the art on the largest availablebenchmarks on Text-based Zero-shot Learning.
展开▼
