Novelty Detection Through Model-Based Characterization of Neural Networks

In this paper, we propose a model-based characterization of neural networks\nto detect novel input types and conditions. Novelty detection is crucial to\nidentify abnormal inputs that can significantly degrade the performance of\nmachine learning algorithms. Majority of existing studies have focused on\nactivation-based representations to detect abnormal inputs, which limits the\ncharacterization of abnormality from a data perspective. However, a model\nperspective can also be informative in terms of the novelties and\nabnormalities. To articulate the significance of the model perspective in\nnovelty detection, we utilize backpropagated gradients. We conduct a\ncomprehensive analysis to compare the representation capability of gradients\nwith that of activation and show that the gradients outperform the activation\nin novel class and condition detection. We validate our approach using four\nimage recognition datasets including MNIST, Fashion-MNIST, CIFAR-10, and\nCURE-TSR. We achieve a significant improvement on all four datasets with an\naverage AUROC of 0.953, 0.918, 0.582, and 0.746, respectively.\n