Reproduce “Preserving Semantic Relations for Zero-Shot Learning” on OpenResearch

About this paper

Yashas Annadani, Soma Biswas

This research introduces a zero-shot learning approach that explicitly preserves semantic relations between classes by mapping class embeddings to a visual feature space. The method achieves state-of-the-art performance on multiple benchmark datasets in both conventional and generalized zero-shot learning settings, demonstrating improved generalization to novel categories and robust recognition.

Key insights

• Explicitly leveraging and preserving the structured relationships within the semantic space is crucial for zero-shot learning models to generalize effectively to unseen categories.
• Mapping semantic information to the visual space can mitigate the hubness problem while still enabling rich semantic representation, provided that semantic relations are meticulously maintained.
• Designing specific objective functions to enforce different types of semantic relations (identical, similar, dissimilar) leads to a more robust and semantically aware embedding space.

Summary by alphaXiv.