Multi-scale Relation Network for Few-Shot Learning Based on Meta-learning

摘要

Deep neural networks can learn a huge function space, because they have millions of parameters to fit large amounts of labeled data. However, this advantage is a major obstacle for few-shot learning, because which has to make predictions based on only few samples of each class. In this work, inspired by multi-scale features methods and relation network which uses neural network to learn metrics, we propose a concise and efficient network, multi-scale relation network. The network consists of a feature extractor and a metric learner. Firstly, the feature extractor extracts multi-scale features by combining features from different convolutional layers. Secondly, we generate the relation feature by calculating the absolute value of the difference between multi-scale features. The results on benchmark sets show that our method avoids the over fitting and elongates the period of learning process, providing higher performance with simple design choices.