Uncertainty quantification is a key pillar of trustworthy machine learning. It enables safe reactions under unsafe inputs, like predicting only when the machine learning model detects sufficient evidence, discarding anomalous data, or emitting warnings when an error is likely to be inbound. This is particularly crucial in safety-critical areas like medical image classification or self-driving cars. Despite the plethora of proposed uncertainty quantification methods achieving increasingly higher scores on performance benchmarks, uncertainty estimates are often shied away from in practice. Many machine learning projects start from pretrained latent representations that come without uncertainty estimates. Uncertainties would need to be trained by practitioners on their own, which is notoriously difficult and resource-intense. This thesis makes uncertainty estimates easily accessible by adding them to the latent representation vectors of pretrained computer vision models. Besides proposing approaches rooted in probability and decision theory, such as Monte-Carlo InfoNCE (MCInfoNCE) and loss prediction, we delve into both theoretical and empirical questions. We show that these unobservable uncertainties about unobservable latent representations are indeed provably correct. We also provide an uncertainty-aware representation learning (URL) benchmark to compare these unobservables against observable ground-truths. Finally, we compile our findings to pretrain lightweight representation uncertainties on large-scale computer vision models that transfer to unseen datasets in a zero-shot manner. Our findings do not only advance the current theoretical understanding of uncertainties over latent variables, but also facilitate the access to uncertainty quantification for future researchers inside and outside the field, enabling straightforward but trustworthy machine learning.
Computational imaging is crucial in many disciplines from autonomous driving to life sciences. However, traditional model-driven and iterative methods consume large computational power and lack scalability for imaging. Deep learning (DL) is effective in processing local-to-local patterns, but it struggles with handling universal global-to-local (nonlocal) patterns under current frameworks. To bridge this gap, we propose a novel DL framework that employs a progressive denoising strategy, named the deterministic diffusion model (DDM), to facilitate general computational imaging at a low cost. We experimentally demonstrate the efficient and faithful image reconstruction capabilities of DDM from nonlocal patterns, such as speckles from multimode fiber and intensity patterns of second harmonic generation, surpassing the capability of previous state-of-the-art DL algorithms. By embedding Bayesian inference into DDM, we establish a theoretical framework and provide experimental proof of its uncertainty quantification. This advancement ensures the predictive reliability of DDM, avoiding misjudgment in high-stakes scenarios. This versatile and integrable DDM framework can readily extend and improve the efficacy of existing DL-based imaging applications.
Uncertainty quantification in travel time estimation (TTE) aims to estimate the confidence interval for travel time, given the origin (O), destination (D), and departure time (T). Accurately quantifying this uncertainty requires generating the most likely path and assessing travel time uncertainty along the path. This involves two main challenges: 1) Predicting a path that aligns with the ground truth, and 2) modeling the impact of travel time in each segment on overall uncertainty under varying conditions. We propose DutyTTE to address these challenges. For the first challenge, we introduce a deep reinforcement learning method to improve alignment between the predicted path and the ground truth, providing more accurate travel time information from road segments to improve TTE. For the second challenge, we propose a mixture of experts guided uncertainty quantification mechanism to better capture travel time uncertainty for each segment under varying contexts. Additionally, we calibrate our results using Hoeffding's upper-confidence bound to provide statistical guarantees for the estimated confidence intervals. Extensive experiments on two real-world datasets demonstrate the superiority of our proposed method.
Uncertainty Quantification in Machine Learning has progressed to predicting the source of uncertainty in a prediction: Uncertainty from stochasticity in the data (aleatoric), or uncertainty from limitations of the model (epistemic). Generally, each uncertainty is evaluated in isolation, but this obscures the fact that they are often not truly disentangled. This work proposes a set of experiments to evaluate disentanglement of aleatoric and epistemic uncertainty, and uses these methods to compare two competing formulations for disentanglement (the Information Theoretic approach, and the Gaussian Logits approach). The results suggest that the Information Theoretic approach gives better disentanglement, but that either predicted source of uncertainty is still largely contaminated by the other for both methods. We conclude that with the current methods for disentangling, aleatoric and epistemic uncertainty are not reliably separated, and we provide a clear set of experimental criteria that good uncertainty disentanglement should follow.
We propose an approximate Bayesian method for quantifying the total uncertainty in inverse PDE solutions obtained with machine learning surrogate models, including operator learning models. The proposed method accounts for uncertainty in the observations and PDE and surrogate models. First, we use the surrogate model to formulate a minimization problem in the reduced space for the maximum a posteriori (MAP) inverse solution. Then, we randomize the MAP objective function and obtain samples of the posterior distribution by minimizing different realizations of the objective function. We test the proposed framework by comparing it with the iterative ensemble smoother and deep ensembling methods for a non-linear diffusion equation with an unknown space-dependent diffusion coefficient. Among other problems, this equation describes groundwater flow in an unconfined aquifer. Depending on the training dataset and ensemble sizes, the proposed method provides similar or more descriptive posteriors of the parameters and states than the iterative ensemble smoother method. Deep ensembling underestimates uncertainty and provides less informative posteriors than the other two methods.
Uncertainty quantification (UQ) is a perspective approach to detecting Large Language Model (LLM) hallucinations and low quality output. In this work, we address one of the challenges of UQ in generation tasks that arises from the conditional dependency between the generation steps of an LLM. We propose to learn this dependency from data. We train a regression model, which target variable is the gap between the conditional and the unconditional generation confidence. During LLM inference, we use this learned conditional dependency model to modulate the uncertainty of the current generation step based on the uncertainty of the previous step. Our experimental evaluation on nine datasets and three LLMs shows that the proposed method is highly effective for uncertainty quantification, achieving substantial improvements over rivaling approaches.
Uncertainty quantification is essential in decision-making, especially when joint distributions of random variables are involved. While conformal prediction provides distribution-free prediction sets with valid coverage guarantees, it traditionally focuses on single predictions. This paper introduces novel conformal prediction methods for estimating the sum or average of unknown labels over specific index sets. We develop conformal prediction intervals for single target to the prediction interval for sum of multiple targets. Under permutation invariant assumptions, we prove the validity of our proposed method. We also apply our algorithms on class average estimation and path cost prediction tasks, and we show that our method outperforms existing conformalized approaches as well as non-conformal approaches.
A neural network architecture is presented that exploits the multilevel properties of high-dimensional parameter-dependent partial differential equations, enabling an efficient approximation of parameter-to-solution maps, rivaling best-in-class methods such as low-rank tensor regression in terms of accuracy and complexity. The neural network is trained with data on adaptively refined finite element meshes, thus reducing data complexity significantly. Error control is achieved by using a reliable finite element a posteriori error estimator, which is also provided as input to the neural network. The proposed U-Net architecture with CNN layers mimics a classical finite element multigrid algorithm. It can be shown that the CNN efficiently approximates all operations required by the solver, including the evaluation of the residual-based error estimator. In the CNN, a culling mask set-up according to the local corrections due to refinement on each mesh level reduces the overall complexity, allowing the network optimization with localized fine-scale finite element data. A complete convergence and complexity analysis is carried out for the adaptive multilevel scheme, which differs in several aspects from previous non-adaptive multilevel CNN. Moreover, numerical experiments with common benchmark problems from Uncertainty Quantification illustrate the practical performance of the architecture.
Complex physical systems are often described by partial differential equations (PDEs) that depend on parameters such as the Reynolds number in fluid mechanics. In applications such as design optimization or uncertainty quantification, solutions of those PDEs need to be evaluated at numerous points in the parameter space. While physics-informed neural networks (PINNs) have emerged as a new strong competitor as a surrogate, their usage in this scenario remains underexplored due to the inherent need for repetitive and time-consuming training. In this paper, we address this problem by proposing a novel extension, parameterized physics-informed neural networks (P$^2$INNs). P$^2$INNs enable modeling the solutions of parameterized PDEs via explicitly encoding a latent representation of PDE parameters. With the extensive empirical evaluation, we demonstrate that P$^2$INNs outperform the baselines both in accuracy and parameter efficiency on benchmark 1D and 2D parameterized PDEs and are also effective in overcoming the known "failure modes".
Accurate classification of haematological cells is critical for diagnosing blood disorders, but presents significant challenges for machine automation owing to the complexity of cell morphology, heterogeneities of biological, pathological, and imaging characteristics, and the imbalance of cell type frequencies. We introduce CytoDiffusion, a diffusion-based classifier that effectively models blood cell morphology, combining accurate classification with robust anomaly detection, resistance to distributional shifts, interpretability, data efficiency, and superhuman uncertainty quantification. Our approach outperforms state-of-the-art discriminative models in anomaly detection (AUC 0.990 vs. 0.918), resistance to domain shifts (85.85% vs. 74.38% balanced accuracy), and performance in low-data regimes (95.88% vs. 94.95% balanced accuracy). Notably, our model generates synthetic blood cell images that are nearly indistinguishable from real images, as demonstrated by an authenticity test in which expert haematologists achieved only 52.3% accuracy (95% CI: [50.5%, 54.2%]) in distinguishing real from generated images. Furthermore, we enhance model explainability through the generation of directly interpretable counterfactual heatmaps. Our comprehensive evaluation framework, encompassing these multiple performance dimensions, establishes a new benchmark for medical image analysis in haematology, ultimately enabling improved diagnostic accuracy in clinical settings. Our code is available at https://github.com/CambridgeCIA/CytoDiffusion.
This work is concerned with conformal prediction in contemporary applications (including generative AI) where a black-box model has been trained on data that are not accessible to the user. Mirroring split-conformal inference, we design a wrapper around a black-box algorithm which calibrates conformity scores. This calibration is local and proceeds in two stages by first adaptively partitioning the predictor space into groups and then calibrating sectionally group by group. Adaptive partitioning (self-grouping) is achieved by fitting a robust regression tree to the conformity scores on the calibration set. This new tree variant is designed in such a way that adding a single new observation does not change the tree fit with overwhelmingly large probability. This add-one-in robustness property allows us to conclude a finite sample group-conditional coverage guarantee, a refinement of the marginal guarantee. In addition, unlike traditional split-conformal inference, adaptive splitting and within-group calibration yields adaptive bands which can stretch and shrink locally. We demonstrate benefits of local tightening on several simulated as well as real examples using non-parametric regression. Finally, we consider two contemporary classification applications for obtaining uncertainty quantification around GPT-4o predictions. We conformalize skin disease diagnoses based on self-reported symptoms as well as predicted states of U.S. legislators based on summaries of their ideology. We demonstrate substantial local tightening of the uncertainty sets while attaining similar marginal coverage.
Data-driven surrogate models have shown immense potential as quick, inexpensive approximations to complex numerical and experimental modelling tasks. However, most surrogate models of physical systems do not quantify their uncertainty, rendering their predictions unreliable, requiring further validation. Though Bayesian approximations offer some solace in estimating the error associated with these models, they cannot provide guarantees, and the quality of their inferences depends on the availability of prior information and good approximations to posteriors for complex problems. This is particularly pertinent to multi-variable or spatio-temporal problems. Our work constructs and formalises a conformal prediction framework that satisfies marginal coverage for spatio-temporal predictions in a model-agnostic manner, requiring near-zero computational costs. We provide an extensive empirical study of the application of the framework to ascertain valid error bars that provide guaranteed coverage across the surrogate model's domain of operation. The application scope of our work extends across a large range of spatio-temporal models, from solving partial differential equations to weather forecasting. Through the applications, the paper looks at providing statistically valid error bars for deterministic models, as well as crafting guarantees to the error bars of probabilistic models. Our conformal prediction formalisation provides guaranteed coverage of the surrogate model, regardless of model architecture, and its training regime and is unbothered by the curse of dimensionality.
In the era of big data, machine learning (ML) has become a powerful tool in various fields, notably impacting structural dynamics. ML algorithms offer advantages by modeling physical phenomena based on data, even in the absence of underlying mechanisms. However, uncertainties such as measurement noise and modeling errors can compromise the reliability of ML predictions, highlighting the need for effective uncertainty awareness to enhance prediction robustness. This paper presents a comprehensive review on navigating uncertainties in ML, categorizing uncertainty-aware approaches into probabilistic methods (including Bayesian and frequentist perspectives) and non-probabilistic methods (such as interval learning and fuzzy learning). Bayesian neural networks, known for their uncertainty quantification and nonlinear mapping capabilities, are emphasized for their superior performance and potential. The review covers various techniques and methodologies for addressing uncertainties in ML, discussing fundamentals and implementation procedures of each method. While providing a concise overview of fundamental concepts, the paper refrains from in-depth critical explanations. Strengths and limitations of each approach are examined, along with their applications in structural dynamic forward problems like response prediction, sensitivity assessment, and reliability analysis, and inverse problems like system identification, model updating, and damage identification. Additionally, the review identifies research gaps and suggests future directions for investigations, aiming to provide comprehensive insights to the research community. By offering an extensive overview of both probabilistic and non-probabilistic approaches, this review aims to assist researchers and practitioners in making informed decisions when utilizing ML techniques to address uncertainties in structural dynamic problems.
When predicting physical phenomena through simulation, quantification of the total uncertainty due to multiple sources is as crucial as making sure the underlying numerical model is accurate. Possible sources include irreducible aleatoric uncertainty due to noise in the data, epistemic uncertainty induced by insufficient data or inadequate parameterization, and model-form uncertainty related to the use of misspecified model equations. Physics-based regularization interacts in nontrivial ways with aleatoric, epistemic and model-form uncertainty and their combination, and a better understanding of this interaction is needed to improve the predictive performance of physics-informed digital twins that operate under real conditions. With a specific focus on biological and physiological models, this study investigates the decomposition of total uncertainty in the estimation of states and parameters of a differential system simulated with MC X-TFC, a new physics-informed approach for uncertainty quantification based on random projections and Monte-Carlo sampling. MC X-TFC is applied to a six-compartment stiff ODE system, the CVSim-6 model, developed in the context of human physiology. The system is analyzed by progressively removing data while estimating an increasing number of parameters and by investigating total uncertainty under model-form misspecification of non-linear resistance in the pulmonary compartment. In particular, we focus on the interaction between the formulation of the discrepancy term and quantification of model-form uncertainty, and show how additional physics can help in the estimation process. The method demonstrates robustness and efficiency in estimating unknown states and parameters, even with limited, sparse, and noisy data. It also offers great flexibility in integrating data with physics for improved estimation, even in cases of model misspecification.
Understanding the mechanisms of interactions within cells, tissues, and organisms is crucial to driving developments across biology and medicine. Mathematical modeling is an essential tool for simulating biological systems and revealing biochemical regulatory mechanisms. Building on experiments, mechanistic models are widely used to describe small-scale intracellular networks and uncover biochemical mechanisms in healthy and diseased states. The rapid development of high-throughput sequencing techniques and computational tools has recently enabled models that span multiple scales, often integrating signaling, gene regulatory, and metabolic networks. These multiscale models enable comprehensive investigations of cellular networks and thus reveal previously unknown disease mechanisms and pharmacological interventions. Here, we review systems biology models from classical mechanistic models to larger, multiscale models that integrate multiple layers of cellular networks. We introduce several examples of models of hypertrophic cardiomyopathy, exercise, and cancer cell proliferation. Additionally, we discuss methods that increase the certainty and accuracy of model predictions. Integrating multiscale models has become a powerful tool for understanding disease and inspiring drug discoveries by incorporating omics data within the cell and across tissues and organisms.
Large climate model ensembles are the primary tool for robustly projecting future climate states and quantifying projection uncertainty. Despite significant advancements in climate modeling over the past few decades, overall projection certainty has not commensurately decreased with steadily improving model skill. We introduce conformal ensembling, a new approach to uncertainty quantification in climate projections based on conformal inference to reduce projection uncertainty. Unlike traditional methods, conformal ensembling seamlessly integrates climate model ensembles and observational data across a range of scales to generate statistically rigorous, easy-to-interpret uncertainty estimates. It can be applied to any climatic variable using any ensemble analysis method and outperforms existing inter-model variability methods in uncertainty quantification across all time horizons and most spatial locations under SSP2-4.5. Conformal ensembling is also computationally efficient, requires minimal assumptions, and is highly robust to the conformity measure. Experiments show that it is effective when conditioning future projections on historical reanalysis data compared with standard ensemble averaging approaches, yielding more physically consistent projections.
Despite the massive advancements in large language models (LLMs), they still suffer from producing plausible but incorrect responses. To improve the reliability of LLMs, recent research has focused on uncertainty quantification to predict whether a response is correct or not. However, most uncertainty quantification methods have been evaluated on single-labeled questions, which removes data uncertainty: the irreducible randomness often present in user queries, which can arise from factors like multiple possible answers. This limitation may cause uncertainty quantification results to be unreliable in practical settings. In this paper, we investigate previous uncertainty quantification methods under the presence of data uncertainty. Our contributions are two-fold: 1) proposing a new Multi-Answer Question Answering dataset, MAQA, consisting of world knowledge, mathematical reasoning, and commonsense reasoning tasks to evaluate uncertainty quantification regarding data uncertainty, and 2) assessing 5 uncertainty quantification methods of diverse white- and black-box LLMs. Our findings show that previous methods relatively struggle compared to single-answer settings, though this varies depending on the task. Moreover, we observe that entropy- and consistency-based methods effectively estimate model uncertainty, even in the presence of data uncertainty. We believe these observations will guide future work on uncertainty quantification in more realistic settings.
Large vision-language models (LVLMs) have made significant progress in recent years. While LVLMs exhibit excellent ability in language understanding, question answering, and conversations of visual inputs, they are prone to producing hallucinations. While several methods are proposed to evaluate the hallucinations in LVLMs, most are reference-based and depend on external tools, which complicates their practical application. To assess the viability of alternative methods, it is critical to understand whether the reference-free approaches, which do not rely on any external tools, can efficiently detect hallucinations. Therefore, we initiate an exploratory study to demonstrate the effectiveness of different reference-free solutions in detecting hallucinations in LVLMs. In particular, we conduct an extensive study on three kinds of techniques: uncertainty-based, consistency-based, and supervised uncertainty quantification methods on four representative LVLMs across two different tasks. The empirical results show that the reference-free approaches are capable of effectively detecting non-factual responses in LVLMs, with the supervised uncertainty quantification method outperforming the others, achieving the best performance across different settings.
Pruning neural networks has proven to be a successful approach to increase the efficiency and reduce the memory storage of deep learning models without compromising performance. Previous literature has shown that it is possible to achieve a sizable reduction in the number of parameters of a deep neural network without deteriorating its predictive capacity in one-shot pruning regimes. Our work builds beyond this background in order to provide rigorous uncertainty quantification for pruning neural networks reliably, which has not been addressed to a great extent in previous literature focusing on pruning methods in computer vision settings. We leverage recent techniques on distribution-free uncertainty quantification to provide finite-sample statistical guarantees to compress deep neural networks, while maintaining high performance. Moreover, this work presents experiments in computer vision tasks to illustrate how uncertainty-aware pruning is a useful approach to deploy sparse neural networks safely.
The success of diffusion probabilistic models in generative tasks, such as text-to-image generation, has motivated the exploration of their application to regression problems commonly encountered in scientific computing and various other domains. In this context, the use of diffusion regression models for ensemble prediction is becoming a practice with increasing popularity. Under such background, we conducted a study to quantitatively evaluate the effectiveness of ensemble methods on solving different regression problems using diffusion models. We consider the ensemble prediction of a diffusion model as a means for zero-shot uncertainty quantification, since the diffusion models in our study are not trained with a loss function containing any uncertainty estimation. Through extensive experiments on 1D and 2D data, we demonstrate that ensemble methods consistently improve model prediction accuracy across various regression tasks. Notably, we observed a larger accuracy gain in auto-regressive prediction compared with point-wise prediction, and that enhancements take place in both the mean-square error and the physics-informed loss. Additionally, we reveal a statistical correlation between ensemble prediction error and ensemble variance, offering insights into balancing computational complexity with prediction accuracy and monitoring prediction confidence in practical applications where the ground truth is unknown. Our study provides a comprehensive view of the utility of diffusion ensembles, serving as a useful reference for practitioners employing diffusion models in regression problem-solving.