- TL;DR: A CNN architecture that can effective rejects the unknowns in test objects
- Abstract: An important property of image classification systems in the real world is that they both accurately classify objects from target classes (``knowns'') and safely reject unknown objects (``unknowns'') that belong to classes not present in the training data. Unfortunately, although the strong generalization ability of existing CNNs ensures their accuracy when classifying known objects, it also causes them to often assign an unknown to a target class with high confidence. As a result, simply using low-confidence detections as a way to detect unknowns does not work well. In this work, we propose an Unknown-aware Deep Neural Network (UDN for short) to solve this challenging problem. The key idea of UDN is to enhance existing CNNs to support a product operation that models the product relationship among the features produced by convolutional layers. This way, missing a single key feature of a target class will greatly reduce the probability of assigning an object to this class. UDN uses a learned ensemble of these product operations, which allows it to balance the contradictory requirements of accurately classifying known objects and correctly rejecting unknowns. To further improve the performance of UDN at detecting unknowns, we propose an information-theoretic regularization strategy that incorporates the objective of rejecting unknowns into the learning process of UDN. We experiment on benchmark image datasets including MNIST, CIFAR-10, CIFAR-100, and SVHN, adding unknowns by injecting one dataset into another. Our results demonstrate that UDN significantly outperforms state-of-the-art methods at rejecting unknowns by 25 percentage points improvement in accuracy, while still preserving the classification accuracy.
- Keywords: unknown, rejection, CNN, product relationship
- Original Pdf: pdf