Uncertainty Quantification in Deep Neural Networks through Statistical Inference on Latent Space

TL;DR

This paper introduces a statistical inference method based on latent space representations to improve uncertainty quantification in deep neural networks, effectively detecting out-of-distribution data and reducing overconfidence.

Contribution

The paper presents a novel algorithm leveraging latent space to assess prediction accuracy and detect outliers, addressing overconfidence issues in existing uncertainty quantification methods.

Findings

Common methods are overconfident on in-distribution data

The proposed method detects out-of-distribution data as inaccurate

The approach improves uncertainty estimation in synthetic datasets

Abstract

Uncertainty-quantification methods are applied to estimate the confidence of deep-neural-networks classifiers over their predictions. However, most widely used methods are known to be overconfident. We address this problem by developing an algorithm that exploits the latent-space representation of data points fed into the network, to assess the accuracy of their prediction. Using the latent-space representation generated by the fraction of training set that the network classifies correctly, we build a statistical model that is able to capture the likelihood of a given prediction. We show on a synthetic dataset that commonly used methods are mostly overconfident. Overconfidence occurs also for predictions made on data points that are outside the distribution that generated the training data. In contrast, our method can detect such out-of-distribution data points as inaccurately predicted, thus aiding in the automatic detection of outliers.