Learning with Analogical Reasoning for Robust Few-Shot Learning

Hailong Zhu, Jun He, Xun Yang, Xueliang Liu, Richang Hong

Keywords: few-shot learning, analogical reasoning, noisy label learning, contrastive learning, transformer

Abstract: Few-shot learning (FSL) is challenging due to limited support data for model training. The situation is much worse when the support data is contaminated with noise. To address this issue, we propose a novel \textbf{T}ransformer-based \textbf{A}nalogical \textbf{R}easoning model for \textbf{N}oisy \textbf{F}ew-\textbf{S}hot learning (TarNFS), by mimicing the human's ability of learning by analogy. Concretely, we assume the existence of a large human cultivated or AI-powered knowledge base, and hypothesize that similar concepts in the knowledge base are visually similar in the latent space as well. Then we design a transformer-based analogical reasoning model to utilize inter-concept connections among these concepts, aiming to build robust and discriminative classification boundaries. In addition, we propose a task-level contrastive learning to analogically learn from negative tasks to facilitate training with noisy tasks. Experiments demonstrate that our TarNFS enables more effective learning from limited and imperfect data. It not only improves the generalization ability of FSL in different noisy settings but also achieves competitive performance in the common clean FSL settings. Code is publicly available \href{https://anonymous.4open.science/r/iclr2088}{here}.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2088