Fine-grained Category Discovery under Coarse-grained supervision with Hierarchical Weighted Self-contrastive Learning

Anonymous

Abstract: In this paper, we propose a new task named Fine-grained Category Discovery under Coarse-grained supervision (FCDC). Without asking for any fine-grained knowledge, FCDC aims at discovering fine-grained categories with only coarse-grained labeled data, which can not only reduce significant labeling costs, but also adapt to novel fine-grained categories. It is also a challenging task since performing FCDC requires models to ensure fine-grained sample separability with only coarse-grained supervision and can easily make models overfit on the training set. Considering most current methods cannot transfer knowledge from coarse-grained level to fine-grained level, we propose a novel hierarchical weighted self-contrastive network to approach the FCDC task. Inspired by the hierarchy of pre-trained models (e.g. BERT), we combine supervised learning and contrastive learning to learn fine-grained knowledge from shallow to deep. Specifically, we use coarse-grained labels to train bottom layers of our model to learn surface knowledge, then we build a novel weighted self-contrastive module to train top layers of our model to learn more fine-grained knowledge. Extensive experiments on two public datasets show both effectiveness and efficiency of our model over state-of-the-art methods.