Meta Learning for Few-Shot One-Class Classification

We propose a method that can perform one-class classification given only small number of examples from the target class and none others. formulate learning meaningful features for as meta-learning problem in which meta-training stage repeatedly simulates classification, using loss chosen algorithm to learn feature representation. To these representations, we require multiclass data similar tasks. show how Support Vector Data Description be used with our method, also simpler variant based on Prototypical Networks obtains comparable performance, indicating representations directly may more important than choose. validate approach by adapting few-shot datasets scenario, obtaining results state-of-the-art traditional improves upon baselines employed setting.