Unsupervised meta-learning for few-shot learning

摘要

Meta-learning is an effective tool to address the few-shot learning problem, which requires new data to be classified considering only a few training examples. However, when used for classification, it requires large labeled datasets, which are not always available in practice. In this paper, we propose an unsupervised meta-learning algorithm that learns from an unlabeled dataset and adapts to downstream human specific tasks with few labeled data. The proposed algorithm constructs tasks using clustering embedding methods and data augmentation functions to satisfy two critical class distinction requirements. To alleviate the biases and the weak diversity problem introduced by data augmentation functions, the proposed algorithm uses two methods, which are shifting the feeding data between the inner-outer loops and a novel data augmentation function. We further provide theoretical analysis of the effect of augmentation data in the inner/outer loop. Experiments on the MiniImagenet and Omniglot datasets demonstrate that the proposed unsupervised meta-learning approach outperforms other tested unsupervised representation learning approaches and two recent unsupervised meta-learning baselines. Compared with supervised meta-learning approaches, certain results produced by our method are quite close to those produced by such methods trained on the human-designed labeled tasks.