Sample-based Uncertainty Quantification with a Single Deterministic Neural Network

TL;DR

This paper improves a neural network-based uncertainty quantification method called DISCO Nets, demonstrating its effectiveness on tabular data, and showing it can produce reliable predictive distributions and feature importance insights.

Contribution

An improved DISCO Nets architecture with faster, more stable training using a compact noise vector, validated on real-world tabular datasets, and a new proof for energy score validity.

Findings

Competitive or superior to standard UQ baselines.

Better point forecast performance than mean squared error trained networks.

Compatible with local feature importance methods like SHAP.

Abstract

Development of an accurate, flexible, and numerically efficient uncertainty quantification (UQ) method is one of fundamental challenges in machine learning. Previously, a UQ method called DISCO Nets has been proposed (Bouchacourt et al., 2016), which trains a neural network by minimizing the energy score. In this method, a random noise vector in $\mathbb{R}^{10\text{--}100}$ is concatenated with the original input vector in order to produce a diverse ensemble forecast despite using a single neural network. While this method has shown promising performance on a hand pose estimation task in computer vision, it remained unexplored whether this method works as nicely for regression on tabular data, and how it competes with more recent advanced UQ methods such as NGBoost. In this paper, we propose an improved neural architecture of DISCO Nets that admits faster and more stable training while only using a compact noise vector of dimension $\sim \mathcal{O}(1)$. We benchmark this approach on miscellaneous real-world tabular datasets and confirm that it is competitive with or even superior to standard UQ baselines. Moreover we observe that it exhibits better point forecast performance than a neural network of the same size trained with the conventional mean squared error. As another advantage of the proposed method, we show that local feature importance computation methods such as SHAP can be easily applied to any subregion of the predictive distribution. A new elementary proof for the validity of using the energy score to learn predictive distributions is also provided.