Reproduce “Absorbing Gradient Conflicts: Modeling Semantic Variance via Kent Distributions for Cross-Modal Hashing” on OpenResearch

About this paper

Hengjie Zhu, Dayan Wu, Zihao Zhang, Xinze Liu, Jingxuan Yu, Peng Fu, Zheng Lin, Weiping Wang

Kent-based Distributional Proxy Hashing (KDPH) models semantic categories using anisotropic Kent distributions on a hypersphere to handle multi-label data in cross-modal hashing. This framework effectively resolves gradient conflicts and prevents proxy collapse, demonstrating improved retrieval performance across benchmark datasets.

Key insights

• Anisotropic Kent distributions can effectively absorb semantic variance and gradient conflicts in multi-label cross-modal hashing by dynamically adjusting their shape parameters rather than just shifting their centroid.
• Modeling class proxies as flexible, anisotropic distributions successfully mitigates 'chaotic proxy oscillation' and prevents 'proxy collapse,' which are common issues in multi-label scenarios with static point proxies.
• The framework's design allows for stable learning of complex probabilistic proxies through geometric constraints and a distribution-aware triplet loss, without incurring additional computational overhead during inference.

Summary by alphaXiv.