Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Distance Awareness

Abstract

Bayesian neural networks (BNN) and deep ensembles are principled approaches to estimate the predictive uncertainty of a deep learning model. However their practicality in real-time, industrial-scale applications are limited due to their heavy memory and inference cost. This motivates us to study principled approaches to high-quality uncertainty estimation that require only a single deep neural network (DNN). By formalizing the uncertainty quantification as a minimax learning problem, we first identify input distance awareness, i.e., the model's ability to quantify the distance of a testing example from the training data in the input space, as a necessary condition for a DNN to achieve high-quality (i.e., minimax optimal) uncertainty estimation. We then propose Spectral-normalized Neural Gaussian Process (SNGP), a simple method that improves the distance-awareness ability of modern DNNs, by adding a weight normalization step during training and replacing the output layer. On a suite of vision and language understanding tasks and on modern architectures (Wide-ResNet and BERT), SNGP is competitive with deep ensembles in prediction, calibration and out-of-domain detection, and outperforms the other single-model approaches.

Explaining Prediction Uncertainty of Pre-trained Language Models by Detecting Uncertain Words in Inputs

“…The problem of predictive uncertainty estimation and calibration has been well studied (Kuleshov and Liang, 2015;Gal and Ghahramani, 2016;Pereyra et al, 2017;Kumar et al, 2018;Liu et al, 2020;Jiang et al, 2021). Calibrated models can make uncertain predictions on some instances, while little is known about what causes prediction uncertainty.…”

Estimating Uncertainty For Vehicle Motion Prediction on Yandex Shifts Dataset

“…We apply spectral-normalized Gaussian process [13] (SNGP) for uncertainty estimation. Since obtaining good representations is crucial for this task, we pretrain plain VectorNet encoder with shallow MLP layer on the same multimodal objective.…”

“…However, the problem of uncertainty estimation (UE) in the context of VMP has not been widely covered yet. Previous works [9], [10], [11] consider a limited number of methods and are either disconnected from state-of-the-art approaches in VMP or recent advances in UE field [12], [13], [14] or both. On the other hand, Bayesian Deep Learning field would benefit from benchmarking on large industrial datasets.…”

DeDUCE: Generating Counterfactual Explanations Efficiently

“…Building on [2], Liu et al [19] show that applying spectral normalisation (SN, [20]) with a coefficient c ≤ 1 to ResNets [9] is enough to enforce both sensitivity and smoothness. Using such models, Mukhoti et al [22] fit a probability distribution to the feature space after the last ResNet block, using the feature representations of the training data.…”