Regularized Meta-Training with Embedding Mixup for Improved Few-Shot Learning

Few-shot learning has enabled techniques to grasp new, unseen tasks from a small set of labelled samples using previously taught knowledge. Although subfields in few-shot learning, such as metric learning, have demonstrated relative success, generalization towards unseen tasks continues to prove difficult, especially in an out-of-domain setting. To address this issue, we propose Embedding Mixup for Meta-Training (EMMeT), a novel regularization technique that creates new tasks through embedding shuffling and averaging for training metric-based backbones. In an experimental setting, our findings across indomain and out-of-domain datasets indicate that application of EMMeT promotes generalization and increases few-shot accuracy across a range of backbone models.