Out of Distribution Data Detection Using Dropout Bayesian Neural Networks

We explore the utility of information contained within a dropout based Bayesian neural network (BNN) for task detecting out distribution (OOD) data. first show how previous attempts to leverage randomized embeddings induced by intermediate layers BNN can fail due distance metric used. introduce an alternative approach measuring embedding uncertainty, and demonstrate incorporating uncertainty improves OOD data identification across three tasks: image classification, language malware detection.