Boosting Few-Shot Learning with Task-Adaptive Multi-level Mixed Supervision

In this paper, we propose a novel task-adaptive few-shot learning (FSL) method called Multi-Level Mixed Supervision (MLMS), which adapts a classifier specifically for each task by mixed supervision. Our method complements the supervised training with a multi-level unsupervised loss including the instance-level certainty term, set-level divergence term, and group-level consistency term. We further modify the set-level divergence term under the unbalanced prior situation where different classes of the unlabeled set contain different numbers of samples. Besides, we propose an approximate solution of minimizing our MLMS loss which is faster than the gradient-based method. Extensive experiments on multiple FSL datasets demonstrate that our method outperforms several recent models by an obvious margin on both transductive FSL and semi-supervised FSL tasks. Codes and trained models are available at https://github.com/Wangduo428/few-shot-learning-mlms.