MetaNODE: Prototype Optimization as a Neural ODE for Few-Shot Learning

Few-Shot Learning (FSL) is a challenging task, i.e., how to recognize novel classes with few examples? Pre-training based methods effectively tackle the problem by pre-training feature extractor and then predicting via cosine nearest neighbor classifier mean-based prototypes. Nevertheless, due data scarcity, prototypes are usually biased. In this paper, we attempt diminish prototype bias regarding it as optimization problem. To end, propose meta-learning framework rectify prototypes, introducing meta-optimizer optimize Although existing meta-optimizers can also be adapted our framework, they all overlook crucial gradient issue, estimation biased on sparse data. address regard its flow meta-knowledge Neural Ordinary Differential Equation (ODE)-based polish called MetaNODE. meta-optimizer, first view initial model process of continuous-time dynamics specified ODE. A inference network carefully designed learn estimate continuous for dynamics. Finally, optimal obtained solving Extensive experiments miniImagenet, tieredImagenet, CUB-200-2011 show effectiveness method.