Cross-modal mapping: Mitigating the modality gap for few-shot classification

doi:10.1016/j.patcog.2026.113285

Cross-modal mapping: Mitigating the modality gap for few-shot classification

2026-02-10

0

PRE

AI

DOI：10.1016/j.patcog.2026.113285

Original

Original request for help

Save

Abstract

En 中文

• A novel Cross-Modal Mapping method is proposed to mitigate the modality gap in vision-language models. • Linear transformation and triplet loss jointly optimize global and local cross-modal alignment. • CMM achieves a 1.06% average accuracy improvement across 11 datasets. • The method demonstrates excellent generalization on distribution shift datasets with high efficiency.

Keywords:

Cross-Modal Mapping

Modality Gap

Few-Shot Classification

Cross-Modal Alignment

Triplet Loss

Cross-modal mapping: Mitigating the modality gap for few-shot classification

Abstract

Journal

Organization