New KDPH framework uses Kent distributions to resolve gradient conflicts in cross-modal hashing

Researchers have introduced Kent-based Distributional Proxy Hashing (KDPH), a novel framework designed to address gradient conflicts in deep cross-modal hashing, particularly in multi-label scenarios. Unlike existing methods that use deterministic points for class proxies, KDPH represents proxies as anisotropic Kent distributions on a hypersphere. This approach allows the model to absorb gradient conflicts by dynamically adjusting its variance, maintaining a stable semantic mean while accommodating diverse label correlations. The framework also incorporates a tailored loss function with a Cayley transform for stable training and orthogonality enforcement. Experiments on benchmark datasets indicate that KDPH effectively mitigates proxy collapse and outperforms current state-of-the-art methods. AI

IMPACT This research could improve the accuracy and stability of large-scale retrieval systems that handle complex, multi-label data.

RANK_REASON The cluster contains an academic paper detailing a new technical framework. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Kent distributions

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →