Bayesian Deep Learning Detection of Anomalies and Failure: Application To Medical Images

Deep learning models trained solely on in-distribution data may not generalize well to anomalous data and may exhibit high confidence in incorrect predictions, leading to significant consequences in safety-critical applications such as medical diagnosis or autonomous driving. It is important for these models to be able to handle anomalous or Out-of-Distribution (OOD) data which are prevalent in many real-world scenarios. In this work, we leverage the predictive variance, i.e., second-order moment of the predictive distribution, of Bayesian models to identify anomalous data points. We test our anomaly and misclassification mechanism on medical image datasets and compare the detection performance of several Bayesian frameworks under various distributional shifts, i.e., noisy conditions, adversarial attacks and OOD data.