FuSiLi: Learning Local Alignment via Global Supervision
irmakbukey · x · 2026-07-18
Introduces **FuSiLi (Fused Sinkhorn-Localized Similarity)**: a multimodal contrastive learning method designed to learn **precise local alignment** even when provided with **only global supervision**. The post uses the analogy of "learning to play along with music by reading sheet music without direct supervision" to illustrate the problem, and includes a link to the paper thread.