Meta learning for few-shot one-class classification

Abstract

Reliably and automatically detecting anomalies is a fundamental problem in several domains. Its applications range from financial security to medical imaging. One common way to address this problem is to frame it as a one-class classification problem: the classifier knows only examples from the normal distribution a priori, and must determine after which of the novel examples are also normal. Despite the success of this approach in some tasks, the recent advances of Machine Learning due to Deep Neural Networks have not yet reached one-class classification techniques. Previous attempts of bringing these advances to the field required compromises, like imposing restrictions to the representational power of the neural networks. This is undesirable because one of the main strengths of the Deep Learning approach is learning useful representations from data directly, instead of relying on manual feature engineering. We propose a method that can perform one-class classification with a different compromise. Our method imposes no restrictions in the network architecture by requiring instead labeled data from related tasks, a requirement which is not available for every scenario. Using these related tasks, we formulate the learning of meaningful features for one- class classification as a meta-learning problem: the meta-training stage repeatedly simulates one-class classification, using the classification loss of the chosen algorithm to learn a feature representation. We show how Support Vector Data Description (SVDD) can be used with our method, and also propose a simpler variant based on Prototypical Networks that obtains comparable performance. This indicates that learning feature representations directly from data may be more important than which one-class algorithm we choose. We validate our approach by adapting few-shot classification datasets to the few-shot one-class classification scenario, obtaining similar results to the state-of-the-art of traditional one-class classification, and that improves upon that of one-class classification baselines employed in the few-shot setting. Moreover, as a practical application, we employ our method to the biometric task of on-device face verification. In this scenario, it compares unfavorably to a standard metric learning technique.