Learning to teach and learn for semi-supervised few-shot image classification

摘要

This paper presents a novel semi-supervised few-shot image classification method named Learning to Teach and Learn (LTTL) to effectively leverage unlabeled samples in small-data regimes. Our method is based on self-training, which assigns pseudo labels to unlabeled data. However, the conventional pseudo-labeling operation heavily relies on the initial model trained by using a handful of labeled data and may produce many noisy labeled samples. We propose to solve the problem with three steps: firstly, cherry-picking searches valuable samples from pseudo-labeled data by using a soft weighting network; and then, cross-teaching allows the classifiers to teach mutually for rejecting more noisy labels. A feature synthesizing strategy is introduced for cross-teaching to avoid clean samples being rejected by mistake; finally, the classifiers are fine-tuned with a few labeled data to avoid gradient drifts. We use the meta-learning paradigm to optimize the parameters in the whole framework. The proposed LTTL combines the power of meta-learning and self-training, achieving superior performance compared with the baseline methods on two public benchmarks.