Résultats de recherche pour « Robust Uncertainty Quantification »

Abstract As Artificial Intelligence (AI) systems are increasingly deployed in high-stakes domains such as healthcare, autonomous systems, finance, and critical infrastructure, ensuring their trustworthiness has become imperative. This paper presents a comprehensive survey of neuro-symbolic AI, a hybrid paradigm that combines the learning capabilities of neural networks with the reasoning strengths of symbolic AI, through the lens of three foundational dimensions: robustness, uncertainty quantification (UQ), and intervenability. We first establish the limitations of purely data-driven “black-box” models in handling distribution shifts, ambiguous inputs, and human oversight. In contrast, neuro-symbolic systems offer enhanced interpretability, verifiability, and control, making them promising candidates for real-world deployment. We systematically review state-of-the-art techniques for modeling robustness, quantifying uncertainty, and enabling intervenability. We further examine how logic, probability, and learning can be integrated into unified or modular architectures to support transparent, adaptive reasoning. Finally, we outline current challenges and identify key research opportunities for advancing neuro-symbolic AI as a trustworthy paradigm. This survey aims to equip researchers and practitioners with a structured understanding of how to build reliable, interpretable, and interactive AI systems by bridging statistical learning and symbolic reasoning.

Reliably quantifying predictive uncertainty is difficult for complex, high-dimensional, or misspecified models. Both fully Bayesian and bootstrap resampling methods provide principled uncertainty estimates but are often too expensive for modern machine-learning models because they require posterior sampling or repeated model refitting. We introduce Ribbon, a scalable approximation to Dirichlet-reweighted bootstrap uncertainty. Ribbon replaces repeated refitting with an influence-function linearization around a single fitted model, preserving the first-order data-reweighting structure of the Bayesian bootstrap while requiring only post-hoc linear algebra. Ribbon approximates the Bayesian-bootstrap or weighted-likelihood-bootstrap refitting target. With a general concentration parameter, Ribbon gives a calibrated Dirichlet-reweighting family whose uncertainty scale can be tuned on validation data. We show that Ribbon is asymptotically equivalent to a flat-prior Laplace approximation under correct likelihood specification and recovers the robust sandwich covariance under misspecification. Across synthetic regression, MNIST classification, and California Housing benchmarks, Ribbon provides competitive predictive performance and improved calibration in several settings while avoiding repeated model retraining.

Uncertainty quantification is critical for deploying deep neural networks (DNNs) in real-world applications. An Auxiliary Uncertainty Estimator (AuxUE) is one of the most effective means to estimate the uncertainty of the main task prediction without modifying the main task model. To be considered robust, an AuxUE must be capable of maintaining its performance and triggering higher uncertainties while encountering Out-of-Distribution (OOD) inputs, i.e., to provide robust aleatoric and epistemic uncertainty. However, for vision regression tasks, current AuxUE designs are mainly adopted for aleatoric uncertainty estimates, and AuxUE robustness has not been explored. In this work, we propose a generalized AuxUE scheme for more robust uncertainty quantification on regression tasks. Concretely, to achieve a more robust aleatoric uncertainty estimation, different distribution assumptions are considered for heteroscedastic noise, and Laplace distribution is finally chosen to approximate the prediction error. For epistemic uncertainty, we propose a novel solution named Discretization-Induced Dirichlet pOsterior (DIDO), which models the Dirichlet posterior on the discretized prediction error. Extensive experiments on age estimation, monocular depth estimation, and super-resolution tasks show that our proposed method can provide robust uncertainty estimates in the face of noisy inputs and that it can be scalable to both image-level and pixel-wise tasks.

Deploying deep learning models in safety-critical applications remains a very challenging task, mandating the provision of assurances for the dependable operation of these models. Uncertainty quantification (UQ) methods estimate the model’s confidence per prediction, informing decision-making by considering the effect of randomness and model misspecification. Despite the advances of state-of-the-art UQ methods, they are computationally expensive or produce conservative prediction sets/intervals. We introduce MC-CP, a novel hybrid UQ method that combines a new adaptive Monte Carlo (MC) dropout method with conformal prediction (CP). MC-CP adaptively modulates the traditional MC dropout at runtime to save memory and computation resources, enabling predictions to be consumed by CP, yielding robust prediction sets/intervals. Throughout comprehensive experiments, we show that MC-CP delivers significant improvements over comparable UQ methods, like MC dropout, RAPS and CQR, both in classification and regression benchmarks. MC-CP can be easily added to existing models, making its deployment simple. The MC-CP code and replication package is available at https://github.com/team-daniel/MC-CP.

The rapid advancement of large language model (LLM) technology has facilitated its integration into various domains of professional and daily life. However, the persistent challenge of LLM hallucination has emerged as a critical limitation, significantly compromising the reliability and trustworthiness of AI-generated content. This challenge has garnered significant attention within the scientific community, prompting extensive research efforts in hallucination detection and mitigation strategies. Current methodological frameworks reveal a critical limitation: traditional uncertainty quantification approaches demonstrate effectiveness primarily within conventional question-answering paradigms, yet exhibit notable deficiencies when confronted with non-canonical or adversarial questioning strategies. This performance gap raises substantial concerns regarding the dependability of LLM responses in real-world applications requiring robust critical thinking capabilities. This study aims to fill this gap by proposing an uncertainty quantification scenario in the task of generating with multiple facts. We have meticulously constructed a set of trap questions contained with fake names. Based on this scenario, we innovatively propose a novel and robust uncertainty quantification method(RU). A series of experiments have been conducted to verify its effectiveness. The results show that the constructed set of trap questions performs excellently. Moreover, when compared with the baseline methods on four different models, our proposed uncertainty quantification method has demonstrated great performance, with an average increase of 0.1-0.2 in ROCAUC values compared to the best performing baseline method, providing new sights and methods for addressing the hallucination issue of LLMs.

Deep learning has become one of the most vital infrastructures under study and entirely a subject of research not at a slow rate that will allow us to comprehend its failures. We apply neural networks to hospitals, financial markets, autonomous systems, and so on, but the culture of the field has traditionally been that benchmark accuracy is the most valued criterion. This book is a corrective. Credible AIs are based on three pillars that are interdependent. Robustness poses whether a model will continue to handle purely outside the regime of its training distribution. Uncertainty quantification inquires a question on whether a model knows which it does not know. Adversarial resilience. The question of adversarial resilience is whether a model can be resilient to its intended manipulation. Each of the challenges is severe in itself; the combination of all of them determines the difference between a model that is functional in the laboratory and one that is worth trusted in the real world. The book is structured in this regard, having three parts, each of them representing the pillar, and a final section, the evaluation frameworks and open problems. The chapters are concluded by exercises as well as annotated literature pointers. The companion repository contains the code of the core examples. It presupposes the knowledge of the basics of deep learning among the readers. It does not need any specific experience in Bayesian statistics or adversarial machine learning. Graduate learners, those who work in the high-stakes fields, and researchers who seek to make cross-cutting in the three areas will all have something appropriate to them. We are honest in our writing of what cannot and can be done by the current method. Strategies that seemed to be good have again and again been defeated by an adaptive attack; distributions that seemed to be estimated have been wrecked by distribution shift. Where evidence is great, we say so. Where it is amalgamated, we say that too. Instead of an optimistic map of the frontier an accurate map of the frontier is of better service to the reader. Deep learning has provided us with unprecedented abilities. There is work to do in order to get those capabilities dependable, truthful about their capabilities and inaccessible to abuse. Hopefully, this book is one of these steps.

The instantaneous reproduction number (${R}_t$) is a key measure of the rate of spread of an infectious disease. Correctly quantifying uncertainty in ${R}_t$ estimates is crucial for making well-informed decisions. Popular ${R}_t$ estimators leverage smoothing techniques to distinguish signal from noise. Examples include EpiEstim and EpiFilter, which are both controlled by a "smoothing parameter" that is traditionally selected by users. We demonstrate that the values of these smoothing parameters are unknown, vary markedly with epidemic dynamics, and show that data-driven smoothing is crucial for accurate uncertainty quantification of real-time ${R}_t$ estimates. We derive novel model likelihoods for the smoothing parameters in both EpiEstim and EpiFilter and develop a Bayesian framework to automatically marginalize these parameters when fitting to epidemiological time-series data. This yields marginal posterior predictive distributions which prove integral to rigorous model evaluation. Applying our methods, we find that default parameterizations of these widely used estimators can negatively impact ${R}_t$ inference, delaying detection of epidemic growth, and misrepresenting uncertainty (typically producing overconfident estimates), with implications for public health decision making. Our extensions mitigate these issues, provide a principled approach to uncertainty quantification, improve the robustness of real-time ${R}_t$ inference, and facilitate model comparison using observable quantities.