The Geometric Mechanics of Contrastive Representation Learning: Alignment Potentials, Entropic Dispersion, and Cross-modal Divergence
对比表示学习的几何力学:对齐势、熵分散和跨模态散度
机构 * University of Science and Technology of China(中国科学技术大学)
AI总结 本文通过测度论框架,在大批量极限下证明InfoNCE目标与确定性能量景观的等价性,揭示单模态与对称多模态之间的几何分岔,并指出跨模态散度项导致模态间隙。
Comments Accepted at ICML 2026; v8: Exposition and notation refined; results unchanged