Few-Shot One-Class Classification via Meta-Learning

Although few-shot learning and one-class classification (OCC), i.e., a binary classifier with data from only one class, have been separately well studied, their intersection remains rather unexplored. Our work addresses the OCC problem presents method to modify episodic sampling strategy of model-agnostic meta-learning (MAML) algorithm learn model initialization particularly suited for tasks. This is done by explicitly optimizing an which requires few gradient steps minibatches yield performance increase on class-balanced test data. We provide theoretical analysis that explains why our approach works in scenario, while other algorithms fail, including unmodified MAML. experiments eight datasets image time-series domains show leads better results than classical approaches, demonstrate ability unseen tasks normal class samples. Moreover, we successfully train anomaly detectors real-world application sensor readings recorded during industrial manufacturing workpieces CNC milling machine, using examples. Finally, empirically proposed technique increases more recent yields state-of-the-art this setting.