Multi-Class Novelty Detection with Generated Hard Novel Features

It has been shown that convolutional neural networks clearly have the overconfidence problem, i.e., mis-classify a novel sample into one of the known classes with high confidence. The task of multi-class novelty detection is thus important to detect novel samples during inference. In this work, we propose to generate hard novel features via a generative adversarial network to facilitate constructing a powerful novelty detector. The generated features should be around the boundaries between known classes and novel classes. They cause a bigger challenge for the novelty detector, and consequently enforce the novelty detector to be stronger. We verify effectiveness of hard novel features from several perspectives, and show that this idea yields the state-of-the-art performance.