Marginally calibrated response distributions for end-to-end learning in autonomous driving

End-to-end learners for autonomous driving are deep neural networks that predict the instantaneous steering angle directly from images of street ahead. These must provide reliable uncertainty estimates their predictions in order to meet safety requirements and initiate a switch manual control areas high uncertainty. However, end-to-end typically only deliver point predictions, since distributional associated with large increases training time or additional computational resources during prediction. To address this shortcoming, we investigate efficient scalable approximate inference model Klein, Nott Smith (J. Comput. Graph. Statist. 30 (2021) 467–483) quantify learners. A special merit model, which refer as implicit copula linear (IC-NLM), is it produces densities marginally calibrated, is, average estimated equals empirical distribution angles. ensure scalability n regimes, develop estimation based on variational fast alternative computationally intensive, exact via Hamiltonian Monte Carlo. We demonstrate accuracy speed approach two trained highway using comma2k19 dataset. The IC-NLM competitive other established quantification methods learning terms nonprobabilistic predictive performance outperforms them marginal calibration in-distribution Our proposed also allows identification overconfident contributes explainability black-box by understand actions learner sees valid.