On Detection of Out of Distribution Inputs in Deep Neural Networks
Susmit Jha, Anirban Roy · 2021
Deep neural networks (DNNs) have achieved near-human-level accuracy on many datasets across different domains. But they are known to produce incorrect predictions with high confidence on out-of-distribution (OOD) inputs. This challenge has limited the adoption of deep learning models in high-assurance systems such as autonomous driving, air traffic man-agement, and medical diagnosis. The problem of detecting when an input is outside the training distribution of a machine learning model, and hence, its prediction on this input cannot be trusted, has received significant attention recently. Several techniques based on statistical, geometric, topological, or relational signature have been developed to detect OOD inputs. In this paper, we investigate two major sources of uncertainty in a deep neural network's prediction on a given input. The first uncertainty source is aleatoric due to the ambiguity or noise in the input. The second is due to epistemic uncertainty arising due to insufficiency of training data and corresponds to the OOD inputs. We posit that the training of deep neural networks using usual objectives cannot distinguish between these two sources of uncertainty. We describe lightweight modifications to the training of deep neural networks that enable deep neural networks to learn representations that can be used to detect OODs. For evaluating our approach, we conducted experiments on CIFAR10 and SVHN as in-distribution data and Imagenet, SVHN (for CIFAR10), and CIFAR10 (for SVHN) as OOD data across different DNN architectures such as ResNet34. WideResNet, and DenseNet.