MASP: Model-Agnostic Sample Propagation for Few-shot learning

Lu Liu, Tianyi Zhou, Guodong Long, Jing Jiang, Xuanyi Dong, Chengqi Zhang

Keywords: few-shot learning, sample propagation, feature calibration, outlier removal, noisy label

Abstract: Few-shot learning aims to train a classifier given only a few samples per class that are highly insufficient to describe the whole data distribution. These few-shot samples not only introduce high variance to the training but also may include outliers near the class boundaries. Directly feeding these samples to training algorithms can lead to unstable optimization and even incorrect gradient descent direction. In this paper, we improve the robustness to ``outliers'' by learning to propagate and refine the representations of few-shot samples to form a more compact data distribution before using them to train a classifier. We develop a mutual calibration among few-shot samples' representations by graph propagation, for which we learn an attention mechanism to build the graph and determine the propagation weights. On both clean datasets and datasets containing noisy labels, we show that our sample propagation generally improves different types of existing few-shot learning methods in multiple few-shot learning settings.

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

One-sentence Summary: We train sample propagation to preprocess few-shot task's data representation before fed into any few-shot learning model, which removes outlier and improves several popular few-shot learning models.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=sVqJpCXOFU