Uncertainty Quantifying Neural Network - Learning Uncertainty Estimates from Monte Carlo Dropout

Quantifying uncertainty in ML models is crucial for ensuring trust in ML models and enhancing safety of ML in critical decision systems. However, many methods of uncertainty quantification are computationally expensive during application. To address this issue, we propose uncertainty quantifying neural networks. The method uniquely combines a two-headed neural network design with pseudo-labels derived from Monte Carlo dropout to efficiently estimate uncertainty while improving computational efficiency. We trained the model’s first head for the given ML task and its second head to predict uncertainty correspondingly. For training the second head, we created pseudo-labels using Monte Carlo dropout. We evaluated the proposed model on different tasks including multi-class classification and regression. We found that the proposed method achieves similar performance on benchmark tasks than pure Monte Carlo dropout with better computational performance during application. Compared to ensemble-based methods, we even achieved superior performance. Theoretically, the proposed model can be adapted to learn uncertainty estimates provided by other methods such as Bayesian neural networks or ensembles. In particular, the high performance with reduced computational resources makes it suitable for time-critical applications.