Estimating Uncertainty with Implicit Quantile Network

TL;DR

This work adapts the Implicit Quantile Network (IQN), originally used for distributional reinforcement learning, to supervised learning by modeling the loss distribution of a trained classifier. By predicting the full loss distribution rather than a scalar loss, the method yields a meaningful uncertainty signal that correlates with incorrect predictions and enables selective rejection of high-uncertainty data, improving accuracy on MNIST and CIFAR without architectural changes. Across experiments, IQN outperforms a scalar-loss baseline and can improve CIFAR100 accuracy by up to roughly 13% when high-uncertainty examples are removed, highlighting its practical value for safety-critical applications. The approach is lightweight compared to ensembles and offers a straightforward path to uncertainty-aware deployment in domains such as healthcare, finance, and autonomous systems.

Abstract

Uncertainty quantification is an important part of many performance critical applications. This paper provides a simple alternative to existing approaches such as ensemble learning and bayesian neural networks. By directly modeling the loss distribution with an Implicit Quantile Network, we get an estimate of how uncertain the model is of its predictions. For experiments with MNIST and CIFAR datasets, the mean of the estimated loss distribution is 2x higher for incorrect predictions. When data with high estimated uncertainty is removed from the test dataset, the accuracy of the model goes up as much as 10%. This method is simple to implement while offering important information to applications where the user has to know when the model could be wrong (e.g. deep learning for healthcare).