Efficient Monte-Carlo Dropout for Uncertainty Quantification

Abstract

Uncertainty quantification is important to realize the confidence of a model's prediction. There are several studied techniques for quantifying the uncertainty of machine learning models one of which is Monte-Carlo dropout. Due to the computational cost of performing inference time modifications to a trained model, Monte Carlo dropouts have rarely been used in performance-oriented machine learning applications. This article discusses a model-splitting technique to modify the Monte-Carlo dropout process to perform uncertainty quantification at the millisecond level. Preliminary analysis shows a potential 25-33 times speedup in performing Monte-Carlo dropouts making it usable for rapid inference.

Keywords: Uncertainty, Machine Learning, Weight, Monte Carlo, Uncertainty Quantification, Efficient, ML, DL