Evidential Deep Learning to Quantify Classification Uncertainty

By Murat Sensoy, Melih Kandemir, Lance M. Kaplan. Neural Information Processing Systems, 2018

1,595 citations recorded by the source corpus.

Deterministic neural nets have been shown to learn effective predictors on a wide range of machine learning problems. However, as the standard approach is to train the network to minimize a prediction loss, the resultant model remains ignorant to its prediction confidence. Orthogonally to Bayesian neural nets that indirectly infer prediction uncertainty through weight uncertainties, we propose explicit modeling of the same using the theory of subjective logic. By placing a Dirichlet distribution on the class probabilities, we treat predictions of a neural net as subjective opinions and learn the function that collects the evidence leading to these opinions by a deterministic neural net from data. The resultant predictor for a multi-class classification problem is another Dirichlet distribution whose parameters are set by the continuous output of a neural net. We provide a preliminary analysis on how the peculiarities of our new loss function drive improved uncertainty estimation. We observe that our method achieves unprecedented success on detection of out-of-distribution queries and endurance against adversarial perturbations.

Subfields: Artificial Intelligence, Safety, Robustness & Interpretability, Probabilistic & Bayesian Methods

View Paper: https://arxiv.org/abs/1806.01768

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 Safety, Robustness & Interpretability

In Probabilistic & Bayesian Methods

Citations in the corpus

Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Distance Awareness (2020, methodology): “We compute the uncertainty score for OOD using the Dempster-Shafer metric as introduced in [68], which empirically leads to better performance for distance-aware models (see Appendix C).”
Being Bayesian, Even Just a Bit, Fixes Overconfidence in ReLU Networks (2020, background): “Malinin & Gales (2018; 2019); Sensoy et al. (2018) used a Dirichlet distribution to model the distribution of a network’s output.”
Towards neural networks that provably know when they don't know (2019, background): “EDL: Evidential deep learning (Sensoy et al., 2018) replaces the softmax layer of a neural network and introduces a different loss function that encourages better uncertainty estimates.”
Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods (2019, background): “Some first proposals for dealing with this case can already be found in the literature (DeVries and Taylor, 2018; Lee et al., 2018b; Malinin and Gales, 2018; Sensoy et al., 2018).”
On Mixup Training: Improved Calibration and Predictive Uncertainty for Deep Neural Networks (2019, background): “For this reason, quantifying the predictive uncertainty and confidence calibration for deep neural networks has seen increased attention in recent years ([6, 18, 7, 9, 19, 15, 28]).”
Trusted Multi-View Classification With Dynamic Evidential Fusion (2022, methodology): “Evidential network [33] models the Dirichlet distribution by introducing subjective logic.”
Drone-Based RGB-Infrared Cross-Modality Vehicle Detection Via Uncertainty-Aware Learning (2020, background): “However, softmax tends to inflate the probability 595 of the predicted class [59], which could affect the detection 596 performance of the cross-modal object detector.”
Explainable AI for Intelligence Augmentation in Multi-Domain Operations (2019, background): “…for the neuro-symbolic hybrid environment, including multi-agent learning with multimodal data (Xing et al. 2018), evidential deep learning (Sensoy, Kaplan, and Kandemir 2018), probabilistic logic programming (Cerutti et al. 2019), forward in-ferencing architectures where the output of a…”
Aleatoric and Epistemic Uncertainty in Machine Learning: A Tutorial Introduction (2019, background): “Some first proposals for dealing with this case can already be found in the literature (DeVries and Taylor, 2018; Lee et al., 2018; Malinin and Gales, 2018; Sensoy et al., 2018).”
Uncertainty-Aware Deep Classifiers Using Generative Models (2020, methodology): “Our work can be considered as an extension of approaches for classiﬁcation, which take the output of a neural network for an input sample to estimate parameters of a Dirichlet distribution (Sensoy, Kaplan, and Kandemir 2018; Malinin and Gales 2018; Gast and Roth 2018) for its classiﬁcation.”
Reverse KL-Divergence Training of Prior Networks: Improved Uncertainty and Adversarial Robustness (2019, neutral): “…of an upper-bound to the cross entropy between discrete distributions, obtained via Jensen’s inequality, which was proposed in a parallel work [34] that investigated a model similar to Dirichlet Prior networks: This form of this upper bound loss is identical to standard negative…”
Theory of Mind and Preference Learning at the Interface of Cognitive Science, Neuroscience, and AI: A Review (2022, background): “Methods from evidential (Sensoy et al., 2018) or epistemic (Cuzzolin, 2021) artiﬁcial intelligence, which model “second order” uncertainty about the probabilities themselves may be interesting to consider to provide a safe and robust approach to ToM in AI.”
A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges (2020, neutral): “[254], Khan et al.”
Uncertainty-aware situational understanding (2019, background): “(10) Let H1 be the hypothesis that the appearance probabilities are equivalent, i.”
Evaluating Scalable Bayesian Deep Learning Methods for Robust Computer Vision (2019, methodology): “For classification tasks, a predictive categorical distribution is commonly realized by a softmax output layer, although recent work has also explored Dirichlet models [14, 41, 34].”
A Review and Comparative Study on Probabilistic Object Detection in Autonomous Driving (2020, methodology): “However, analytically calculating the posterior distribution is intractable due to its high dimensionality and multi-modality (often in a space of millions of weight parameters), as well as the non-linear activation function between consecutive layers of deep neural networks [23].”
A survey of uncertainty in deep neural networks (2021, methodology): “the training process [44] or external components that have to be trained and evaluated additionally [46].”