ODIN: Outlier Detection In Neural Networks

Rickard Sjögren, Johan Trygg

Abstract: Adoption of deep learning in safety-critical systems raise the need for understanding what deep neural networks do not understand. Several methodologies to estimate model uncertainty have been proposed, but these methodologies constrain either how the neural network is trained or constructed. We present Outlier Detection In Neural networks (ODIN), an assumption-free method for detecting outlier observations during prediction, based on principles widely used in manufacturing process monitoring. By using a linear approximation of the hidden layer manifold, we add prediction-time outlier detection to models after training without altering architecture or training. We demonstrate that ODIN efficiently detect outliers during prediction on Fashion-MNIST, ImageNet-synsets and speech command recognition.

Keywords: Outlier Detection, Model Uncertainty, Safety

TL;DR: An add-on method for deep learning to detect outliers during prediction-time