Task-Independent Knowledge Makes for Transferable Representations for Generalized Zero-Shot Learning

Generalized Zero-Shot Learning (GZSL) targets recognizing new categories by learning transferable image representations. Existing methods find that, aligning representations with corresponding semantic labels, the semantic-aligned can be transferred to unseen categories. However, supervised only seen category learned knowledge is highly task-specific, which makes biased towards In this paper, we propose a novel Dual-Contrastive Embedding Network (DCEN) that simultaneously learns task-specific and task-independent via alignment instance discrimination. First, DCEN leverages task labels cluster of same cross-modal contrastive exploring semantic-visual complementarity. Besides knowledge, then introduces attracting different views repelling images. Compared high-level supervision, discrimination supervision encourages capture low-level visual less toward alleviates representation bias. Consequently, jointly make for DCEN, obtains averaged 4.1% improvement on four public benchmarks.