Use of Uncertainty with Autoencoder Neural Networks for Anomaly Detection

Autoencoders neural networks are nonlinear dimension reduction models widely used in the field of anomaly detection. Conventionally, the reconstruction error is considered as a score function allowing the discrimination between the normal data and the outliers. Recent advances in calculating uncertainty from neural networks open new perspectives in the field of anomaly detection. We study, for given models and different concentrations of anomalies, several score functions. We compare the standard score function based on the standard error, a score based on the error resulting from the Bayesian approximation, as well as score functions directly including the uncertainty. This paper empirically demonstrates how including uncertainty in the score function is likely to improve the performance of an autoencoder-based anomaly detection model.