Abstract:A popular strategy for active learning is to specifically target a reduction in epistemic uncertainty, since aleatoric uncertainty is often considered as being intrinsic to the system of interest and therefore not reducible. Yet, distinguishing these two types of uncertainty remains challenging and there is no single strategy that consistently outperforms the others. We propose to use a particular combination of probability and possibility theories, with the aim of using the latter to specifically represent epistemic uncertainty, and we show how this combination leads to new active learning strategies that have desirable properties. In order to demonstrate the efficiency of these strategies in non-trivial settings, we introduce the notion of a possibilistic Gaussian process (GP) and consider GP-based multiclass and binary classification problems, for which the proposed methods display a strong performance for both simulated and real datasets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to improve the handling of epistemic uncertainty (EU) in active learning, thereby enhancing the learning efficiency and performance of the model. ### Problem Background In active learning, choosing which unlabeled data points to label in order to minimize the required amount of data and obtain an efficient model is a crucial issue. Traditional active learning strategies usually aim to reduce epistemic uncertainty because it is caused by a lack of knowledge, while aleatoric uncertainty (AU) is considered an inherent and non - reducible characteristic of the system. However, distinguishing between these two types of uncertainty remains a challenge, and no single strategy can always outperform others. ### Core Contributions of the Paper 1. **Introducing a New Representation Method of Uncertainty**: - The authors propose a method that combines probability theory and possibility theory, using the latter to specifically represent epistemic uncertainty. - In this way, they develop new active learning strategies that have desirable properties. 2. **Introducing the Possibilistic Gaussian Process (PGP)**: - To verify the effectiveness of these strategies in non - trivial settings, the authors introduce the concept of the Possibilistic Gaussian Process. - PGP is a new framework for active learning in multi - class and binary classification problems. 3. **Performance Evaluation**: - The authors demonstrate the strong performance of the proposed method on simulated and real - world datasets, where at least one of the proposed acquisition functions outperforms the baseline method in each case. ### Formula Representation Some of the key formulas involved in the paper are as follows: - **Possibility Function**: \[ N(\theta; \mu, \Sigma) = \exp\left(-\frac{1}{2} (\theta - \mu)^T \Sigma^{-1} (\theta - \mu)\right) \] This is the multivariate Gaussian possibility function, where \(\mu\) is the mean vector and \(\Sigma\) is the covariance matrix. - **Conditional Possibility Function**: \[ f_{\theta_1|\theta_2}(\theta_1 | \theta_2) = N(\theta_1; \mu_{1|2}, \Sigma_{1|2}) \] where, \[ \mu_{1|2} = \mu_1 + \Sigma_{21} \Sigma_{22}^{-1} (\theta_2 - \mu_2) \] \[ \Sigma_{1|2} = \Sigma_{11} - \Sigma_{12} \Sigma_{22}^{-1} \Sigma_{21} \] - **Epistemic Uncertainty Measure**: \[ U_\Theta^{f_\theta} = \int_\Theta f_\theta(\theta) d\theta \] \[ U_Y^{\bar{P}} = \int_Y \sup_{\theta \in \Theta} \left(f_\theta(\theta) p_Y(y | \theta)\right) dy - 1 \] ### Summary This paper provides a more effective active learning method to handle epistemic uncertainty by introducing a new possibility theory framework and PGP. The experimental results show that this method performs well on multiple tasks, especially in classification problems.

Improving Active Learning with a Bayesian Representation of Epistemic Uncertainty

Uncertainty-Based Active Learning Via Sparse Modeling for Image Classification

Evidential uncertainty sampling for active learning

Robustness Guarantees for Bayesian Inference with Gaussian Processes

What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

Double-Bayesian Learning

Bayesian Layers: A Module for Neural Network Uncertainty

Bayesian probabilistic propagation of imprecise probabilities with large epistemic uncertainty

Bayesian Surrogate Analysis and Uncertainty Propagation

The Epistemic Uncertainty Hole: an issue of Bayesian Neural Networks

Query Efficient Posterior Estimation in Scientific Experiments via Bayesian Active Learning

Bayesian Learning of Orthogonal Embeddings for Multi-Fidelity Gaussian Processes

Iterative Construction of Gaussian Process Surrogate Models for Bayesian Inference

Fixed-Mean Gaussian Processes for Post-hoc Bayesian Deep Learning

Optimal simulation-based Bayesian decisions

Generative Posterior Networks for Approximately Bayesian Epistemic Uncertainty Estimation

Near-Optimal Bayesian Active Learning with Noisy Observations

Active Learning with Fully Bayesian Neural Networks for Discontinuous and Nonstationary Data

On Uncertainty Quantification for Near-Bayes Optimal Algorithms

Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?

Looking at the posterior: accuracy and uncertainty of neural-network predictions