Uncertainty In Artificial Intelligence Research Articles

Can input reconstruction be used to directly estimate uncertainty of a dose prediction U-Net model?

The reliable and efficient estimation of uncertainty in artificial intelligence (AI) models poses an ongoing challenge in many fields such as radiation therapy. AI models are intended to automate manual steps involved in the treatment planning workflow. We focus in this study on dose prediction models that predict an optimal dose trade-off for each new patient for a specific treatment modality. They can guide physicians in the optimization, be part of automatic treatment plan generation or support decision in treatmentindication. Most common uncertainty estimation methods are based on Bayesian approximations, like Monte Carlo dropout (MCDO) or Deep ensembling (DE). These two techniques, however, have a high inference time (i.e., require multiple inference passes) and might not work for detecting out-of-distribution (OOD) data (i.e., overlapping uncertainty estimate for in-distribution (ID) and OOD). In this study, we present a direct uncertainty estimation method and apply it for a dose prediction U-Net architecture. It can be used to flag OOD data and give information on the quality of the doseprediction. Our method consists in the addition of a branch decoding from the bottleneck which reconstructs the CT scan given as input. The input reconstruction error can be used as a surrogate of the model uncertainty. For the proof-of-concept, our method is applied to proton therapy dose prediction in head and neck cancer patients. A dataset of 60 oropharyngeal patients was used to train the network using a nested cross-validation approach with 11 folds (training: 50 patients, validation: 5 patients, test: 5 patients). For the OOD experiment, we used 10 extra patients with a different head and neck sub-location. Accuracy, time-gain, and OOD detection are analyzed for our method in this particular application and compared with the popular MCDO andDE. The additional branch did not reduce the accuracy of the dose prediction model. The median absolute error is close to zero for the target volumes and less than 1% of the dose prescription for organs at risk. Our input reconstruction method showed a higher Pearson correlation coefficient with the prediction error (0.620) than DE (0.447) and MCDO (between 0.599 and 0.612). Moreover, our method allows an easier identification of OOD (no overlap for ID and OOD data and a Z-score of 34.05). The uncertainty is estimated simultaneously to the regression task, therefore requires less time and computationalresources. This study shows that the error in the CT scan reconstruction can be used as a surrogate of the uncertainty of the model. The Pearson correlation coefficient with the dose prediction error is slightly higher than state-of-the-art techniques. OOD data can be more easily detected and the uncertainty metric is computed simultaneously to the regression task, therefore faster than MCDO or DE. The code and pretrained model are available on the gitlab repository: https://gitlab.com/ai4miro/ct-reconstruction-for-uncertainty-quatification-of-hdunet.

Communication of Uncertainty in AI Regulations

Scholarship of uncertainty in artificial intelligence (AI) regulation has focused on theories, strategies, and practices to mitigate uncertainty. However, there is little understanding of how federal agencies communicate scientific uncertainties to all stakeholders including the public and regulated industries. This is important for three reasons: one, it highlights what aspects of the issue are quantifiable; two, it displays how agencies explain uncertainties about the issues that are not easily quantified; and three, it shows how knowledgeable agencies perceive the public audience in relation to the issue at hand and what they expect from such communication. By analyzing AI regulations across four categories of scientific uncertainties, this study found that uncertainty in areas of ownership, safety, and transparency are hard to quantify and hence agencies use personalized examples to explain uncertainties. In addition, agencies seek public input to gather additional data and derive consensus on issues that have moral implications. These findings are consistent with the literature on tackling uncertainty and regulatory decision-making. They can help advance our understanding of current practices of communicating science effectively to explain risks and uncertainties.

Communication of Uncertainty in AI Regulations

Scholarship of uncertainty in artificial intelligence (AI) regulation has focused on theories, strategies, and practices to mitigate uncertainty. However, there is little understanding of how federal agencies communicate scientific uncertainties to all stakeholders including the public and regulated industries. This is important for three reasons: one, it highlights what aspects of the issue are quantifiable; two, it displays how agencies explain uncertainties about the issues that are not easily quantified; and three, it shows how knowledgeable agencies perceive the public audience in relation to the issue at hand and what they expect from such communication. By analyzing AI regulations across four categories of scientific uncertainties, this study found that uncertainty in areas of ownership, safety, and transparency are hard to quantify and hence agencies use personalized examples to explain uncertainties. In addition, agencies seek public input to gather additional data and derive consensus on issues that have moral implications. These findings are consistent with the literature on tackling uncertainty and regulatory decision-making. They can help advance our understanding of current practices of communicating science effectively to explain risks and uncertainties.

Estimation of separable direct and indirect effects in a continuous-time illness-death model.

In this article we study the effect of a baseline exposure on a terminal time-to-event outcome either directly or mediated by the illness state of a continuous-time illness-death process with baseline covariates. We propose a definition of the corresponding direct and indirect effects using the concept of separable (interventionist) effects (Robins and Richardson in Causality and psychopathology: finding the determinants of disorders and their cures, Oxford University Press, 2011; Robins et al. in arXiv:2008.06019 , 2021; Stensrud et al. in J Am Stat Assoc 117:175-183, 2022). Our proposal generalizes Martinussen and Stensrud (Biometrics 79:127-139, 2023) who consider similar causal estimands for disentangling the causal treatment effects on the event of interest and competing events in the standard continuous-time competing risk model. Unlike natural direct and indirect effects (Robins and Greenland in Epidemiology 3:143-155, 1992; Pearl in Proceedings of the seventeenth conference on uncertainty in artificial intelligence, Morgan Kaufmann, 2001) which are usually defined through manipulations of the mediator independently of the exposure (so-called cross-world interventions), separable direct and indirect effects are defined through interventions on different components of the exposure that exert their effects through distinct causal mechanisms. This approach allows us to define meaningful mediation targets even though the mediating event is truncated by the terminal event. We present the conditions for identifiability, which include some arguably restrictive structural assumptions on the treatment mechanism, and discuss when such assumptions are valid. The identifying functionals are used to construct plug-in estimators for the separable direct and indirect effects. We also present multiply robust and asymptotically efficient estimators based on the efficient influence functions. We verify the theoretical properties of the estimators in a simulation study, and we demonstrate the use of the estimators using data from a Danish registry study.

Estimating diagnostic uncertainty in artificial intelligence assisted pathology using conformal prediction

Unreliable predictions can occur when an artificial intelligence (AI) system is presented with data it has not been exposed to during training. We demonstrate the use of conformal prediction to detect unreliable predictions, using histopathological diagnosis and grading of prostate biopsies as example. We digitized 7788 prostate biopsies from 1192 men in the STHLM3 diagnostic study, used for training, and 3059 biopsies from 676 men used for testing. With conformal prediction, 1 in 794 (0.1%) predictions is incorrect for cancer diagnosis (compared to 14 errors [2%] without conformal prediction) while 175 (22%) of the predictions are flagged as unreliable when the AI-system is presented with new data from the same lab and scanner that it was trained on. Conformal prediction could with small samples (N = 49 for external scanner, N = 10 for external lab and scanner, and N = 12 for external lab, scanner and pathology assessment) detect systematic differences in external data leading to worse predictive performance. The AI-system with conformal prediction commits 3 (2%) errors for cancer detection in cases of atypical prostate tissue compared to 44 (25%) without conformal prediction, while the system flags 143 (80%) unreliable predictions. We conclude that conformal prediction can increase patient safety of AI-systems.

The need for quantification of uncertainty in artificial intelligence for clinical data analysis: increasing the level of trust in the decision-making process

Different terms such as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">trust</i> , <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">certainty</i> , and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">uncertainty</i> are of great importance in the real world and play a critical role in artificial intelligence (AI) applications. The implied assumption is that the level of trust in AI can be measured in different ways. This principle can be achieved by distinguishing uncertainties in predicting AI methods used in medical studies. Hence, it is necessary to propose effective uncertainty quantification (UQ) and measurement methods to have trustworthy AI (TAI) clinical decision support systems (CDSSs). In this study, we present practical guidelines for developing and using UQ methods while applying various AI techniques for medical data analysis.

Local Explanations via Necessity and Sufficiency: Unifying Theory and Practice

Necessity and sufficiency are the building blocks of all successful explanations. Yet despite their importance, these notions have been conceptually underdeveloped and inconsistently applied in explainable artificial intelligence (XAI), a fast-growing research area that is so far lacking in firm theoretical foundations. In this article, an expanded version of a paper originally presented at the 37th Conference on Uncertainty in Artificial Intelligence (Watson et al., 2021), we attempt to fill this gap. Building on work in logic, probability, and causality, we establish the central role of necessity and sufficiency in XAI, unifying seemingly disparate methods in a single formal framework. We propose a novel formulation of these concepts, and demonstrate its advantages over leading alternatives. We present a sound and complete algorithm for computing explanatory factors with respect to a given context and set of agentive preferences, allowing users to identify necessary and sufficient conditions for desired outcomes at minimal cost. Experiments on real and simulated data confirm our method’s competitive performance against state of the art XAI tools on a diverse array of tasks.

Selection and integration of generalized instrumental variables for estimating total effects

We consider a situation where cause–effect relationships between variables can be described as a directed acyclic graph (DAG) and the corresponding linear structural equation model (linear SEM). When several pairs of instrumental variables (IVs) and covariates (IV-pairs; Pearl in: Proceedings of the 20th conference on uncertainty in artificial intelligence, AUAI Press, Arlington, Virginia, United States, UAI’04, 2004) are available, we propose (i) the graphical selection criteria of IV-pairs for total effects and (ii) an integrated estimator that combines them to estimate total effects with better accuracy. In this paper, in accordance with the paper by Brito and Pearl (in: Proceedings of the 18th conference on uncertainty in artificial intelligence, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, UAI’02, 2002), the proposed estimator is called an integrated generalized instrumental variable (iGIV) estimator. The proposed graphical selection criteria imply that (a) the back-door criterion achieves better estimation accuracy than the traditional instrumental variable (IV) method of estimating total effects even when the treatment and covariates are highly correlated and (b) the conditional IV method can be superior to the back-door criterion in some situations. The iGIV estimator provides a general class that includes both the ordinary least squares (OLS) estimator based on the back-door criterion and the two-stage least squares (2SLS) estimator based on the (conditional) IV method. We clarify the properties of the iGIV estimator, some of which can be read off from the DAG structure. Furthermore, through numerical experiments and an application to a case study, we show that the performance of the iGIV estimator is superior to those of the OLS and IV estimators. The iGIV estimator can be a powerful tool to estimate the total effect when the proposed graphical selection criteria of the IV-pairs are not satisfied.

Using fuzzy measures for modeling human perception of uncertainty in artificial intelligence

We observe the importance of modeling human perceptions of various types of information for the construction of A. I. systems and highlight the need for modeling human perceptions of uncertainty. We introduce the concept of a fuzzy measure, μ, and discuss its use as a general structure for the representation of knowledge about an uncertain variable. We note the ability of the fuzzy measure to model various formulations of uncertain information in a unified format. We look at various properties of fuzzy, particularly the ability to fuse multiple fuzzy measures to form new measures. We introduce various operations on fuzzy measures motivated from probability theory such as the determination of expected values and variances. We emphasize the importance of the Choquet integral in these operations. We discuss two characterizing features of a fuzzy measure, its entropy and attitudinal character, and note their usefulness in helping select an appropriate measure. We finally look at the issue of answering questions about variables having information about its value modeled by a fuzzy measure. Here we must come to grips with the fact that for most measures a fundamental property of the probability measure, Prob(A) + Prob(A¯) = 1, does not hold.

De novo gene signature identification from single-cell RNA-seq with hierarchical Poisson factorization.

Common approaches to gene signature discovery in single‐cell RNA‐sequencing (scRNA‐seq) depend upon predefined structures like clusters or pseudo‐temporal order, require prior normalization, or do not account for the sparsity of single‐cell data. We present single‐cell hierarchical Poisson factorization (scHPF), a Bayesian factorization method that adapts hierarchical Poisson factorization (Gopalan et al, 2015, Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence, 326) for de novo discovery of both continuous and discrete expression patterns from scRNA‐seq. scHPF does not require prior normalization and captures statistical properties of single‐cell data better than other methods in benchmark datasets. Applied to scRNA‐seq of the core and margin of a high‐grade glioma, scHPF uncovers marked differences in the abundance of glioma subpopulations across tumor regions and regionally associated expression biases within glioma subpopulations. scHFP revealed an expression signature that was spatially biased toward the glioma‐infiltrated margins and associated with inferior survival in glioblastoma.

Margins of discrete Bayesian networks

Bayesian network models with latent variables are widely used in statistics and machine learning. In this paper, we provide a complete algebraic characterization of these models when the observed variables are discrete and no assumption is made about the state-space of the latent variables. We show that it is algebraically equivalent to the so-called nested Markov model, meaning that the two are the same up to inequality constraints on the joint probabilities. In particular, these two models have the same dimension, differing only by inequality constraints for which there is no general description. The nested Markov model is therefore the closest possible description of the latent variable model that avoids consideration of inequalities. A consequence of this is that the constraint finding algorithm of Tian and Pearl [In Proceedings of the 18th Conference on Uncertainty in Artificial Intelligence (2002) 519–527] is complete for finding equality constraints. Latent variable models suffer from difficulties of unidentifiable parameters and nonregular asymptotics; in contrast the nested Markov model is fully identifiable, represents a curved exponential family of known dimension, and can easily be fitted using an explicit parameterization.

Deep bidirectional intelligence: AlphaZero, deep IA-search, deep IA-infer, and TPC causal learning

This paper starts at a brief review on AlphaGoZero, Q learning, and Monte-Carlo tree search (MCTS), in a comparison with decades ago studies on A* search and CNneim-A that was proposed in 1986 and shares a scouting technique similar to one used in MCTS. Then, we combine the strengths of AlphaGoZero and CNneim-A, resulting in a family named deep IA-search that consists of Deep Scout A*, Deep CNneim-A, Deep Bi-Scout A*, and V-AlphaGoZero, as well as their extensions. Moreover, relation between search and reasoning motivates to extend deep IA-search to Deep IA-Infer for implementing reasoning. Especially, another early study (Xu and Pearl, Structuring causal tree models with continuous variables. In: Proceedings of the 3rd annual conference on uncertainty in artificial intelligence, pp 170–179 1987) on structuring causal tree is developed into a three phase causal learning approach, namely topology identification, parameter reestimation, and causal rho-tree search on a casual rho-diagram that is defined by a set of pairwise correlation coefficient rho. Algorithms are sketched for discovering casual topologies of triplets, stars, and trees, as well as some topologies of casual rho-Directed Acyclic Graph (DAG), e.g. ones for Yule–Simpson’s paradox, Pearl’s Sprinkler DAG, and Back door DAG. Furthermore, the classic Boolean SAT problem is extended into one rho-SAT problem, and the roles of four fundamental mechanisms in an intelligent system are elaborated, with insights on integrating these mechanisms to encode not only variables but also how they are organised, as well as on why deep networks are preferred while extra depth is unnecessary.

Reasoning with alternative acyclic directed mixed graphs

Acyclic directed mixed graphs (ADMGs) are the graphs used by Pearl (Causality: models, reasoning, and inference. Cambridge University Press, Cambridge, 2009) for causal effect identification. Recently, alternative acyclic directed mixed graphs (aADMGs) have been proposed by Peña (Proceedings of the 32nd conference on uncertainty in artificial intelligence, 577–586, 2016) for causal effect identification in domains with additive noise. Since the ADMG and the aADMG of the domain at hand may encode different model assumptions, it may be that the causal effect of interest is identifiable in one but not in the other. Causal effect identification in ADMGs is well understood. In this paper, we introduce a sound algorithm for identifying arbitrary causal effects from aADMGs. We show that the algorithm follows from a calculus similar to Pearl’s do-calculus. Then, we turn our attention to Andersson–Madigan–Perlman chain graphs, which are a subclass of aADMGs, and propose a factorization for the positive discrete probability distributions that are Markovian with respect to these chain graphs. We also develop an algorithm to perform maximum likelihood estimation of the factors in the factorization.

Causal feature learning: an overview

Causal feature learning (CFL) (Chalupka et al., Proceedings of the Thirty-First Conference on Uncertainty in Artificial Intelligence. AUAI Press, Edinburgh, pp 181–190, 2015) is a causal inference framework rooted in the language of causal graphical models (Pearl J, Reasoning and inference. Cambridge University Press, Cambridge, 2009; Spirtes et al., Causation, Prediction, and Search. Massachusetts Institute of Technology, Massachusetts, 2000), and computational mechanics (Shalizi, PhD thesis, University of Wisconsin at Madison, 2001). CFL is aimed at discovering high-level causal relations from low-level data, and at reducing the experimental effort to understand confounding among the high-level variables. We first review the scientific motivation for CFL, then present a detailed introduction to the framework, laying out the definitions and algorithmic steps. A simple example illustrates the techniques involved in the learning steps and provides visual intuition. Finally, we discuss the limitations of the current framework and list a number of open problems.

Interpretive Strategies for Screen-Based Creative Technologies

This paper brings together the disciplines of media and creative technologies studies and software systems engineering; it focuses on the challenge of finding methodologies to measure, test and decode meaning in digital cultural objects. Just as rough set theory is a mathematical tool to deal with vagueness and uncertainty in artificial intelligence, and approximation accuracy and knowledge granularity are approaches to uncertainty research, the authors argue that découpage analytique is a possible method for decoding screen-based information. They draw on a variety of examples: interactive online digital art projects; an interactive, immersive screen-based art installation; re-mediated digital art installation; expanded cinema; a videogame; and a medical interface example, in order to determine if it is possible to map interpretive strategies that include a blending of old and new criteria, but ultimately promoting an equal partnership between artist and audience, and thus, a community of co-creators. Additionally, the authors present experimental evidence on the difference introduced by the screen size to further qualify the effectiveness of découpage analytique in relation to the amount of screen real estate afforded.

Mutual Authentication Technique with Four Entities Implemented by Fuzzy Neural Network in 4-G Mobile Communications

G mobile communications system is offering high speed data communications technology having connectivity to all sorts of the networks including 2-G and 3-G mobile networks. Authentication of a mobile subscriber (MS) or a sub network and a main network are an important issue to check and minimize security threats or attacks. An advanced artificial intelligence based mutual authentication technique executed by fuzzy neural network with four entities is proposed. Voice frequency of the salutation or the selective words used by a subscriber like Hello, Good Morning, etc. is taken as first entity. Second entity is chosen as thumb fingerprint matching of the calling subscriber with his / her stored thumb fingerprint. Then third entity is taken as face image matching of the calling subscriber. Fourth entity is granted as probability of the salutation word from subscribers talking habit while initializing a call. These four entities such as probability of particular range of frequencies for the salutation word, the thumb fingerprint matching, the face image matching of the subscriber, using particular salutation or greeting word at the time of starting a call are used with most frequently, more frequently and less frequently by the calling subscriber like uncertainty in Artificial Intelligence. Now different relative grades are assigned to most frequently, more frequently and less frequently used parameters. Fuzzy operations such as intersection and union are computed taking three membership functions at a time out of four membership functions to adopt fuzzy neural network. Thereafter, the optimum or the final fuzzy operations are computed according to the assumed weightages. Lastly, the optimized fuzzy operations are defuzzified by the Composite Maxima method, and the results are tested according to the invented fuzzy neural rule. If the results are satisfactory, the subscriber or the sub network and the network (the

Using Bayesian networks for highly available cloud-based web applications

Bayesian networks have demonstrated their capability in several applications spanning from reasoning under uncertainty in artificial intelligence to dependability modelling and analysis. This paper focuses on the use of this language for allocating cloud resources to maximise service dependability. This objective is accomplished by the definition of a model-driven approach able to guide the software engineering to define a cloud infrastructure (applications, services, virtual and concrete resources) using a semi-automated process. This process exploits both high-level languages such as UML as well as Bayesian networks. Using all their features (backward analysis, ease of usage, low analysis time), Bayesian networks are used in this process as a driver for the optimization, learning and estimation phases. The paper discusses all the issues that the application of Bayesian networks in the proposed process arises.

On the Read-Once Property of Branching Programs and CNFs of Bounded Treewidth

In this paper we prove a space lower bound of $$n^{\varOmega (k)}$$ for non-deterministic (syntactic) read-once branching programs (nrobps) on functions expressible as cnfs with treewidth at most k of their primal graphs. This lower bound rules out the possibility of fixed-parameter space complexity of nrobps parameterized by k. We use lower bound for nrobps to obtain a quasi-polynomial separation between Free Binary Decision Diagrams and Decision Decomposable Negation Normal Forms, essentially matching the existing upper bound introduced by Beame et al. (Proceedings of the twenty-ninth conference on uncertainty in artificial intelligence, Bellevue, 2013) and thus proving the tightness of the latter.

Frontiers for propositional reasoning about fragments of probabilistic conditional independence and hierarchical database decompositions

Conditional independence provides an essential framework to deal with knowledge and uncertainty in Artificial Intelligence, and is fundamental in probability and multivariate statistics. Its associated implication problem is paramount for building Bayesian networks. Unfortunately, the problem does not enjoy an axiomatization, by a finite set of Horn rules, and is already coNP-complete to decide for some fragments of conditional independencies. Saturated conditional independencies form an important fragment whose implication problem is decidable in almost linear time. We establish an axiomatization, by a finite set of Horn rules, for the fragment of generalized saturated conditional independencies. These state the conditional independence between finitely many sets of random variables. The special case of two sets captures Geiger and Pearl's finite axiomatization for saturated conditional independencies. Even for this special case, our completeness proof is new. The proof argument utilizes special probability models where two events have probability one half. Special probability models allow us to establish an equivalence between the implication of generalized saturated conditional independencies and that of formulae in a Boolean propositional fragment. This duality is then extended to a trinity including the implication problem of Delobel's full first-order hierarchical database decompositions. Already for the independence between two sets of random variables, the dualities cannot be extended to cover conditional independencies in general, or first-order hierarchical decompositions.

Causal discovery in an adult ADHD data set suggests indirect link between DAT1 genetic variants and striatal brain activation during reward processing.

Attention-deficit/hyperactivity disorder (ADHD) is a common and highly heritable disorder affecting both children and adults. One of the candidate genes for ADHD is DAT1, encoding the dopamine transporter. In an attempt to clarify its mode of action, we assessed brain activity during the reward anticipation phase of the Monetary Incentive Delay (MID) task in a functional MRI paradigm in 87 adult participants with ADHD and 77 controls (average age 36.5 years). The MID task activates the ventral striatum, where DAT1 is most highly expressed. A previous analysis based on standard statistical techniques did not show any significant dependencies between a variant in the DAT1 gene and brain activation [Hoogman et al. (2013); Neuropsychopharm 23:469-478]. Here, we used an alternative method for analyzing the data, that is, causal modeling. The Bayesian Constraint-based Causal Discovery (BCCD) algorithm [Claassen and Heskes (2012); Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence] is able to find direct and indirect dependencies between variables, determines the strength of the dependencies, and provides a graphical visualization to interpret the results. Through BCCD one gets an opportunity to consider several variables together and to infer causal relations between them. Application of the BCCD algorithm confirmed that there is no evidence of a direct link between DAT1 genetic variability and brain activation, but suggested an indirect link mediated through inattention symptoms and diagnostic status of ADHD. Our finding of an indirect link of DAT1 with striatal activity during reward anticipation might explain existing discrepancies in the current literature. Further experiments should confirm this hypothesis. © 2015 Wiley Periodicals, Inc.