UQGAN: A Unified Model for Uncertainty Quantification of Deep Classifiers trained via Conditional GANs

Abstract

We present an approach to quantifying both aleatoric and epistemic uncertainty for deep neural networks in image classification, based on generative adversarial networks (GANs). While most works in the literature that use GANs to generate out-of-distribution (OoD) examples only focus on the evaluation of OoD detection, we present a GAN based approach to learn a classifier that exhibits proper uncertainties for OoD examples as well as for false positives (FPs). Instead of shielding the entire in-distribution data with GAN generated OoD examples which is state-of-the-art, we shield each class separately with out-of-class examples generated by a conditional GAN and complement this with a one-vs-all image classifier. In our experiments, in particular on CIFAR10, we improve over the OoD detection and FP detection performance of state-of-the-art GAN-training based classifiers. Furthermore, we also find that the generated GAN examples do not significantly affect the calibration error of our classifier and result in a significant gain in model accuracy.

A Survey on Uncertainty Quantification Methods for Deep Neural Networks: An Uncertainty Source Perspective

“…the flow model is integrated into GAN to enable likelihood estimation and better uncertainty quantification. Another approach [39,99] further extends the method by utilizing the Wasserstein GAN [4] based on gradient penalty to improve model convergence. The key advantage of deep generative modeling for uncertainty quantification is that they directly parameterize a deep neural network to represent the prediction distribution, without the need to have an explicit distribution format.…”

“…One direction of uncertainty quantification for PINN is to build a probabilistic neural network 𝑝 𝜽 (𝒖|𝒙, 𝑡, 𝒛) for propagating uncertainty originating from various sources, where 𝒖 (𝒙, 𝑡) represents the spatio-temporal random field with the explanatory spatial variable 𝒙 and temporal variable 𝑡. 𝒛 is a latent variable that follows a prior distribution 𝑝 (𝒛) and aim to encode the variability of the high-dimensional observation 𝒖 into low-dimensional embedding. many approaches based on the deep generative model for uncertainty quantification have been proposed to leverage their capability of probabilistic modeling for structured outputs [26,39,99].…”