Pitfalls of Epistemic Uncertainty Quantification through Loss Minimisation
This work highlights a critical flaw in uncertainty quantification methods for machine learning practitioners, making it incremental by identifying pitfalls in existing proposals.
The paper analyzes a recent second-order learner approach for quantifying epistemic uncertainty and shows that loss minimization fails to incentivize faithful representation of epistemic uncertainty.
Uncertainty quantification has received increasing attention in machine learning in the recent past. In particular, a distinction between aleatoric and epistemic uncertainty has been found useful in this regard. The latter refers to the learner's (lack of) knowledge and appears to be especially difficult to measure and quantify. In this paper, we analyse a recent proposal based on the idea of a second-order learner, which yields predictions in the form of distributions over probability distributions. While standard (first-order) learners can be trained to predict accurate probabilities, namely by minimising suitable loss functions on sample data, we show that loss minimisation does not work for second-order predictors: The loss functions proposed for inducing such predictors do not incentivise the learner to represent its epistemic uncertainty in a faithful way.