Evaluation of Neural Network Uncertainty Estimation with Application to Resource-Constrained Platforms

The ability to accurately estimate uncertainties in neural network predictions is of great importance in many critical tasks. In this paper, we first analyze the intrinsic relation between two main use cases of uncertainty estimation, i.e., selective prediction and confidence calibration. We then reveal the potential issues with the existing quality metrics for uncertainty estimation and propose new metrics to mitigate them. Finally, we apply these new metrics to resource-constrained platforms such as autonomous driver assistance systems where the quality of uncertainty estimation is critical. By exploring the trade-off between the model size and the estimation quality, a missing piece in the literature, some interesting trends are observed.