Prototype Reinforcement for Few-Shot Learning

Few-shot learning requires to recognize novel classes with scarce labeled data. The effectiveness of Prototypical Networks has been recognized in existing studies, however, training on the narrow-size distribution of scarce data usually tends to get biased prototypes. In this paper, we figure out two influencing factors of the process: the feature redundancy and the feature monotony. We propose a simple but effective method to reinforce the prototype. In our method, feature transformation and feature drift are used to reduce the feature redundancy of the prototype. Besides, the pseudo-label strategy is used to recalculate the prototype, enrich the features of the prototype, and alleviate the problem of feature monotony. Effectiveness is shown on two few-shot benchmarks, mini-ImageNet, and CUB. The experimental results show that compared with the ordinary mean prototype, the reinforcement prototype can effectively improve the classification accuracy.