Direct Training of Dynamic Observation Noise with UMarineNet

Accurate uncertainty predictions are crucial to assess the reliability of a model, especially for neural networks. Part of this uncertainty is the observation noise, which is dynamic in our marine virtual sensor task. Typically, dynamic noise is not trained directly, but approximated through terms in the loss function. Unfortunately, this noise loss function needs to be scaled by a trade-off-parameter to achieve accurate uncertainties. In this paper we propose an upgrade to the existing architecture, which increases interpretability and introduces a novel direct training procedure for dynamic noise modelling. To that end, we train the point prediction model and the noise model separately. We present a new loss function that requires Monte Carlo runs of the model to directly train for the uncertainty prediction accuracy. In an experimental evaluation, we show that in most tested cases the uncertainty prediction is more accurate than the manually tuned trade-off-parameter. Because of the architectural changes we are able to analyze the importance of individual parts of the time series of our prediction.