Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning

By Arsenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, Dmitry P. Vetrov. International Conference on Learning Representations, 2020

362 citations recorded by the source corpus.

Uncertainty estimation and ensembling methods go hand-in-hand. Uncertainty estimation is one of the main benchmarks for assessment of ensembling performance. At the same time, deep learning ensembles have provided state-of-the-art results in uncertainty estimation. In this work, we focus on in-domain uncertainty for image classification. We explore the standards for its quantification and point out pitfalls of existing metrics. Avoiding these pitfalls, we perform a broad study of different ensembling techniques. To provide more insight in the broad comparison, we introduce the deep ensemble equivalent (DEE) and show that many sophisticated ensembling techniques are equivalent to an ensemble of very few independently trained networks in terms of the test log-likelihood.

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/d12fd94337ac804470dc78911e74b5b6480eef8e

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

A high-resolution canopy height model of the Earth (2022, methodology): “It can be interpreted as a natural variant of test-time augmentation, which has been demonstrated to improve the calibration of uncertainty estimates from deep ensembles in the domain of computer vision 37 .”
A survey of uncertainty in deep neural networks (2021, background): “• In-domain uncertainty [65] In-domain uncertainty represents the uncertainty related to an input drawn from a data distribution assumed to be equal to the training data distribution.”
MEMO: Test Time Robustness via Adaptation and Augmentation (2021, methodology): “Both TTA and MEMO are also effective at improving performance for the original CIFAR-10 test set where there is no distribution shift, providing further support for the widespread use of augmentations in standard evaluation protocols [19, 1].”
What Are Bayesian Neural Network Posteriors Really Like? (2021, background): “Multiple works have considered tempering the posterior in Bayesian neural networks (e.g. Wenzel et al., 2020; Wilson & Izmailov, 2020; Zhang et al., 2020b; Ashukha et al., 2020; Aitchison, 2020).”
Machine learning and deep learning—A review for ecologists (2022, methodology): “New methods such as Bayesian neural networks have paved the way to obtain uncertainty and prediction intervals for DL models, bringing ML algorithms closer to statistical models (Ashukha et al., 2021 ; Loquercio et al., 2020 ).”
Efficient Test-Time Model Adaptation without Forgetting (2022, methodology): “Actually, one can address this by either maintaining a sliding window { x i } ti = t − L including L previous samples and the current test sample x t , or data augmentation techniques adopted in existing single sample TTA methods ((Zhang et al., 2021b; Khurana et al., 2021)).”
Second opinion needed: communicating uncertainty in medical machine learning (2021, methodology): “For neural networks and deep learning methods, some simple methods to calculate conditional uncertainty estimates include Monte Carlo (MC) Dropout 29 and ensembling 30 – 32 .”
Few-shot Conformal Prediction with Auxiliary Tasks (2021, background): “Unlike CP, however, these methods are not guaranteed to be accurate, and often suffer from miscalibration in practice—and this is especially true for modern neural networks (Guo et al., 2017; Ashukha et al., 2020; Hirschfeld et al., 2020).”
Class-Incremental Domain Adaptation with Smoothing and Calibration for Surgical Report Generation (2021, methodology): “We evaluate the model using four metrics for image captioning, namely BLEUn [20], ROUGE [15], METEOR [3], CIDEr [23] and miscalibration with Expected Calibration Error (ECE), Static Calibration Error (SCE), Thresholded Adaptive Calibration Error (TACE) [where, threshold = 10−3], Brier Score (BS) [2,”
A Review and Comparative Study on Probabilistic Object Detection in Autonomous Driving (2020, methodology): “MC-Dropout on the other hand has been shown to signiﬁcantly under-perform ensembles and other methods on simple predictive uncertainty estimation tasks such as classiﬁcation in [64], [114].”
Uncertainty Estimation with Infinitesimal Jackknife, Its Distribution and Mean-Field Approximation (2020, background): “This ensemble of models can be obtained by varying training with respect to initialization [28], hyper-parameters [3], data partitions, i.e ., bootstraps [14, 42].”
Uncertainty Estimation in Autoregressive Structured Prediction (2021, background): “Thirdly, we are unable to determine whether the error is due to knowledge or An in-depth comparison of ensemble methods was conducted in (Ashukha et al., 2020; Ovadia et al., 2019) Data and Knowledge Uncertainty are sometimes also called Aleatoric and Epistemic uncertainty.”
Bayesian Deep Learning and a Probabilistic Perspective of Generalization (2020, background): “, 2019; Ashukha et al., 2020) show that deep ensembles appear to outperform some particular approaches to Bayesian neural networks for uncertainty representation, leading to the confusion that deep ensembles and Bayesian methods are competing approaches. These methods are often explicitly referred t”
Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty (2020, background): “Currently, deep ensembles [102] are one of best performing uncertainty quantification approaches for NNs [8], retaining calibration even under dataset shift [141].”
A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges (2020, methodology): “[198] Image Neural Networks Deep ensemble equivalent score (DEE) L(w) = − 1 N ∑N i=1 log p̂(y ? i |xi, w)+ λ 2 ‖w‖2 →”