Prototype and Metric Based Prediction for Data-Efficient Training

We propose a prototype- and metric-based prediction method together with several training pipelines suitable for training a network without using any additional data in the few-shot learning tasks with different intra-class variances. Being tested on two datasets commonly used for few-shot learning, our method has shown satisfactory ability to improve data efficiency and prevent overfitting. It even competes with the meta-learning-based method trained with a lot of extra labeled samples on the dataset with low intra-class variance and shows no significant performance gap when it comes to the dataset with a high intra-class variance. We reported 99.0% acc on the Omniglot dataset and 48.0% acc on the mini-ImageNet for 5-way 5-shot tasks.