Metric Learning for Novelty and Anomaly Detection
This addresses the issue of unreliable predictions in real-world applications like traffic sign recognition, but it is incremental as it builds on existing out-of-distribution detection work.
The paper tackles the problem of neural networks making overconfident errors on out-of-distribution images by proposing metric learning to improve novelty and anomaly detection, achieving comparable or better results than previous methods in experiments including traffic sign recognition.
When neural networks process images which do not resemble the distribution seen during training, so called out-of-distribution images, they often make wrong predictions, and do so too confidently. The capability to detect out-of-distribution images is therefore crucial for many real-world applications. We divide out-of-distribution detection between novelty detection ---images of classes which are not in the training set but are related to those---, and anomaly detection ---images with classes which are unrelated to the training set. By related we mean they contain the same type of objects, like digits in MNIST and SVHN. Most existing work has focused on anomaly detection, and has addressed this problem considering networks trained with the cross-entropy loss. Differently from them, we propose to use metric learning which does not have the drawback of the softmax layer (inherent to cross-entropy methods), which forces the network to divide its prediction power over the learned classes. We perform extensive experiments and evaluate both novelty and anomaly detection, even in a relevant application such as traffic sign recognition, obtaining comparable or better results than previous works.