Regularization with Multiple Feature Combination for Few-Shot Learning

Few-shot learning solves problems with a limited amount of labeled examples. Our analysis shows the existing metric-based methods concentrate on highly discriminative features while not fully utilizing whole capacity. In this work, we propose a novel regularization technique that constrains the model to exploit whole capacity by distinguishing data with multiple feature combinations. Our approach achieves state-of the-art performance in several public benchmarks compared to the existing metric-based methods.