Layer-Wise Adaptive Updating for Few-Shot Image Classification

Few-shot image classification (FSIC), which requires a model to recognize new categories via learning from few images of these categories, has attracted lots of attention. Recently, meta-learning based methods have been shown is promising for FSIC. Commonly, they train a meta-learner to learn easy fine-tuning weight for FSIC. When solving an FSIC task, the meta-learner efficiently updates itself on few images of the task and turns to a task-specific model. In this letter, we propose a novel meta-learning based layer-wise adaptive updating (LWAU) method for FSIC. LWAU is inspired by an interesting finding that compared with common deep models, the meta-learners pay much more attention to update their top layers when learning from few images. According to this finding, we assume that the meta-learners may greatly prefer updating their top layers to updating their bottom layers for better FSIC performance. Therefore, in LWAU, the meta-learner is trained to learn not only the easy fine-tuning weight but also its favorite layer-wise adaptive updating rule. Extensive experiments show that compared with existing few-shot classification methods, the proposed LWAU: 1) achieves better FSIC performance with a clear margin; 2) almost updates only its top layer when solving FSIC, which indicates the learned feature extractor is much more generalizable; 3) learns sparser feature extractor; 4) learns from few images more efficiently by at least 5 times.