Uncertainty Estimation for Neural Time Series with an Application to Sideslip Angle Estimation

The automotive industry offers many applications for machine learning (ML), in general, and deep neural networks in particular. However, the real-world deployment of neural networks into safety-critical components remains a challenge as models would need to offer robustness under a wide range of operating conditions. In this work, we focus on uncertainty estimation, which can be used to deliver predictors that fail gracefully, by detecting situations where their predictions are unreliable. Following Gräber et al. [1], we use Recurrent Neural Networks (RNNs) to perform sideslip angle estimation. To perform robust uncertainty estimation, we augment the RNNs with generative models. We demonstrate the advantage of the proposed model architecture over Monte Carlo (MC) dropout [2] on the Revs data set [3].