An Optimized Model based on Metric-Learning for Few-Shot Classification

Abstract

Few-shot learning makes up for the shortcomings of traditional deep learning that requires a large amount of labeled data, and has great potential in promoting machines to become more intelligent. Many existing few-shot learning methods have achieved benign performance in numerous classification tasks by training a classifier, yet some trained models are restricted by shallow networks which will gravely restrict their feature expression ability. In addition, what proves awful is that some previous few-shot learning methods do not use appropriate loss functions to train excellent models, which limits their performance to some extent. To settle above problems, we optimize the classical few-shot learning framework, that is, prototypical networks, from three aspects: data augmentation, increasing the network's feature expression ability and improving the training loss function. It is worth mentioning that besides keeping simple and efficient, our innovative metric-learning-based few-shot classification framework is capable to be integrated into the same model to achieve end-to-end training. Immense amounts of experimental results show that our model not only performs well in classification tasks, but also shows its amazing superiority and competitiveness compared with related technologies.