Noise-Aware Few-Shot Learning through Bi-directional Multi-view Prompt Alignment

ICLR 2026 Conference Submission16067 Authors

Keywords: Robust Learning, Learning with Noisy Label, Few-shot Learning, Vision-Language Model

TL;DR: Noise-Aware Few-Shot Learning through Bi-directional Multi-view Prompt Alignment

Abstract: Vision-language models offer strong few-shot capability through prompt tuning but remain vulnerable to noisy labels, which can corrupt prompts and degrade cross-modal alignment. Existing approaches struggle because they often lack the ability to model fine-grained semantic cues and to adaptively separate clean from noisy signals. To address these challenges, we propose NA-MVP, a framework for **N**oise-**A**ware few-shot learning through bi-directional **M**ulti-**V**iew **P**rompt alignment. NA-MVP introduces three key innovations: multi-view prompts with unbalanced optimal transport alignment that enable fine-grained patch-to-prompt matching while suppressing unreliable regions and reinforcing clean correspondences; a bi-directional prompt design that jointly models clean-oriented and noise-aware semantics to disentangle useful signals from corrupted ones; and adaptive sample refinement with optimal transport that employs a learnable threshold to correct mislabeled samples while preserving reliable data. Experiments on both synthetic and real-world noisy datasets demonstrate that NA-MVP consistently outperforms state-of-the-art baselines, highlighting its effectiveness for robust few-shot learning under noisy supervision.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16067