Theoretical Refinement of CLIP by Utilizing Linear Structure of Optimal Similarity

Name: Theoretical Refinement of CLIP by Utilizing Linear Structure of Optimal Similarity
Uploaded: 2026-10-04T02:09:59.801681+00:00
Duration: 659 s
Description: KME-CLIP enhances CLIP's similarity computation by exploiting the linear structure of Pointwise Mutual Information (PMI) within a Reproducing Kernel Hilbert Space (RKHS). The method theoretically proves that this approach can approximate PMI with arbitrary accuracy, leading to empirically superior performance in retrieval and classification tasks across various benchmarks.

KME-CLIP enhances CLIP's similarity computation by exploiting the linear structure of Pointwise Mutual Information (PMI) within a Reproducing Kernel Hilbert Space (RKHS). The method theoretically proves that this approach can approximate PMI with arbitrary accuracy, leading to empirically superior performance in retrieval and classification tasks across various benchmarks.

Open the full episode