Abstract
To endow neural networks with the potential to learn rapidly from a handful of samples, meta-learning blazes a trail to acquire across-task knowledge from a variety of few-shot learning tasks. However, most existing meta-learning algorithms retain the requirement of fine-grained supervision, which is expensive in many applications. In this paper, we show that meta-learning models can extract transferable knowledge from coarse-grained supervision for few-shot classification. We propose a weakly-supervised framework, namely Coarse-to-fine Pseudo-labeling Guided Meta-Learning (CPGML), to alleviate the need for data annotation. In our framework, the coarse-categories are grouped into pseudo sub-categories to construct a task distribution for meta-training, following the cosine distance between the corresponding embedding vectors of images. For better feature representation in this process, we develop Dual-level Discriminative Embedding (DDE) aiming to keep the distance between learned embeddings consistent with the visual similarity and semantic relation of input images simultaneously. Moreover, we propose a task-attention mechanism to reduce the weight of the training tasks with potentially higher label noises based on the observation of task-nonequivalence. Extensive experiments conducted on two hierarchical meta-learning benchmarks demonstrate that, under the proposed framework, meta-learning models can effectively extract task-independent knowledge from the roughly-generated tasks and generalize well to unseen tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2007.05675
