An Adaptive Plug-and-Play Network for Few-Shot Learning (2302.09326v1)

Abstract: Few-shot learning (FSL) requires a model to classify new samples after learning from only a few samples. While remarkable results are achieved in existing methods, the performance of embedding and metrics determines the upper limit of classification accuracy in FSL. The bottleneck is that deep networks and complex metrics tend to induce overfitting in FSL, making it difficult to further improve the performance. Towards this, we propose plug-and-play model-adaptive resizer (MAR) and adaptive similarity metric (ASM) without any other losses. MAR retains high-resolution details to alleviate the overfitting problem caused by data scarcity, and ASM decouples the relationship between different metrics and then fuses them into an advanced one. Extensive experiments show that the proposed method could boost existing methods on two standard dataset and a fine-grained datasets, and achieve state-of-the-art results on mini-ImageNet and tiered-ImageNet.