FEMTO-ST researchers boost neural network uncertainty estimates with Dirichlet MC Dropout

Traditional neural networks output deterministic predictions with no inherent uncertainty, a critical limitation for high-stakes applications like medical diagnosis or autonomous driving. While Bayesian Neural Networks (BNNs) provide principled uncertainty quantification, they are computationally expensive and hard to scale. Monte Carlo (MC) Dropout has emerged as an efficient approximation to Bayesian inference by performing multiple stochastic forward passes, but its uncertainty estimates can be noisy.

In this new work, Rouaa Hoblos and colleagues at FEMTO-ST enhance MC Dropout by integrating a Dirichlet distribution over class probabilities, following the framework of Sensoy et al. (2018). This Dirichlet-based modeling captures richer, more informative uncertainty representations—distinguishing between epistemic (model) and aleatoric (data) uncertainty. The key advantage is that the method retains the low computational overhead of standard MC Dropout while significantly improving the calibration and reliability of uncertainty estimates. The paper details theoretical foundations, comparisons with existing techniques, and experimental results confirming the effectiveness of the approach. This provides a practical, lightweight solution for adding robust uncertainty awareness to deep neural networks.