Christophe Grojean, Ayan Paul, Zhuoni Qian et al.
Adding interpretability to multivariate methods creates a powerful synergy for exploring complex physical systems with higher order correlations while bringing about a degree of clarity in the underlying dynamics of the system.
Steven Hicks, Andrea Storås, Michael Riegler et al.
Deep learning has in recent years achieved immense success in all areas of computer vision and has the potential of assisting medical doctors in analyzing visual content for disease and other abnormalities. However, the current state of deep learning is very much a black box, making medical professionals highly skeptical about integrating these methods into clinical practice. Several methods have been proposed in order to shine some light onto these black boxes, but there is no consensus on the opinion of the medical doctors that will consume these explanations. This paper presents a study asking medical doctors about their opinion of current state-of-the-art explainable artificial intelligence methods when applied to a gastrointestinal disease detection use case. We compare two different categories of explanation methods, intrinsic and extrinsic, and gauge their opinion of the current value of these explanations. The results indicate that intrinsic explanations are preferred and that explanation.
Patrik Hammersborg, Inga Strümke
With large chess-playing neural network models like AlphaZero contesting the state of the art within the world of computerised chess, two challenges present themselves: The question of how to explain the domain knowledge internalised by such models, and the problem that such models are not made openly available. This work presents the re-implementation of the concept detection methodology applied to AlphaZero in McGrath et al. (2022), by using large, open-source chess models with comparable performance. We obtain results similar to those achieved on AlphaZero, while relying solely on open-source resources. We also present a novel explainable AI (XAI) method, which is guaranteed to highlight exhaustively and exclusively the information used by the explained model. This method generates visual explanations tailored to domains characterised by discrete input spaces, as is the case for chess. Our presented method has the desirable property of controlling the information flow between any input vector and the given model, which in turn provides strict guarantees regarding what information is used by the trained model during inference. We demonstrate the viability of our method by applying it to standard 8x8 chess, using large open-source chess models.
Andrea M. Storås, Anders Åsberg, Pål Halvorsen et al.
Tacrolimus is one of the cornerstone immunosuppressive drugs in most transplantation centers worldwide following solid organ transplantation. Therapeutic drug monitoring of tacrolimus is necessary in order to avoid rejection of the transplanted organ or severe side effects. However, finding the right dose for a given patient is challenging, even for experienced clinicians. Consequently, a tool that can accurately estimate the drug exposure for individual dose adaptions would be of high clinical value. In this work, we propose a new technique using machine learning to estimate the tacrolimus exposure in kidney transplant recipients. Our models achieve predictive errors that are at the same level as an established population pharmacokinetic model, but are faster to develop and require less knowledge about the pharmacokinetic properties of the drug.
Vilde B. Gjærum, Inga Strümke, Ole Andreas Alsos et al.
Deep neural networks (DNNs) can be useful within the marine robotics field, but their utility value is restricted by their black-box nature. Explainable artificial intelligence methods attempt to understand how such black-boxes make their decisions. In this work, linear model trees (LMTs) are used to approximate the DNN controlling an autonomous surface vessel (ASV) in a simulated environment and then run in parallel with the DNN to give explanations in the form of feature attributions in real-time. How well a model can be understood depends not only on the explanation itself, but also on how well it is presented and adapted to the receiver of said explanation. Different end-users may need both different types of explanations, as well as different representations of these. The main contributions of this work are (1) significantly improving both the accuracy and the build time of a greedy approach for building LMTs by introducing ordering of features in the splitting of the tree, (2) giving an overview of the characteristics of the seafarer/operator and the developer as two different end-users of the agent and receiver of the explanations, and (3) suggesting a visualization of the docking agent, the environment, and the feature attributions given by the LMT for when the developer is the end-user of the system, and another visualization for when the seafarer or operator is the end-user, based on their different characteristics.
Michael A. Riegler, Inga Strümke
We present swarm-attack, an open-source adversarial testing framework in which multiple lightweight LLM agents coordinate through shared memory, parallel exploration, and evolutionary optimization. Together, our results demonstrate that both safety bypass of frontier models and software vulnerability discovery, i.e., the capability class that motivated restricted release of Anthropic's Mythos Preview, are achievable at effectively zero cost using commodity hardware and openly available models. We report two experiments. In the first, five instances of a 1.2 billion parameter model conducted 225 jailbreak attacks each against GPT-4o and Claude Sonnet~4. Against GPT-4o, the swarm achieved an Effective Harm Rate of 45.8%, producing 49 critical-severity breaches; against Claude Sonnet-4, the Effective Harm Rate was 0% despite a 40% technical success rate. In the second experiment, the same models performed combined source code analysis and binary fuzzing against a vulnerable C application with 9 planted CWEs. With a hand-crafted exploit seed corpus, regex pattern detection, and AddressSanitizer-based crash classification, the pipeline recovers 9 of 9 vulnerabilities (100% recall) in approximately four minutes on a consumer MacBook. With those scaffold components disabled, the same model recovers 0 of 9 by crash verification and 2 of 9 by citation. The capability class that motivated restricted release of Anthropic's Mythos Preview is therefore reproducible at effectively zero cost; the important enabler is the system scaffold itself, which compensates for the limited reasoning capacity of small individual models.
Inga Strümke, Marija Slavkovik, Clemens Stachl
Decisions such as which movie to watch next, which song to listen to, or which product to buy online, are increasingly influenced by recommender systems and user models that incorporate information on users' past behaviours, preferences, and digitally created content. Machine learning models that enable recommendations and that are trained on user data may unintentionally leverage information on human characteristics that are considered vulnerabilities, such as depression, young age, or gambling addiction. The use of algorithmic decisions based on latent vulnerable state representations could be considered manipulative and could have a deteriorating impact on the condition of vulnerable individuals. In this paper, we are concerned with the problem of machine learning models inadvertently modelling vulnerabilities, and want to raise awareness for this issue to be considered in legislation and AI ethics. Hence, we define and describe common vulnerabilities, and illustrate cases where they are likely to play a role in algorithmic decision-making. We propose a set of requirements for methods to detect the potential for vulnerability modelling, detect whether vulnerable groups are treated differently by a model, and detect whether a model has created an internal representation of vulnerability. We conclude that explainable artificial intelligence methods may be necessary for detecting vulnerability exploitation by machine learning-based recommendation systems.
Inga Strümke, Marija Slavkovik
If a prediction model identifies vulnerable individuals or groups, the use of that model may become an ethical issue. But can we know that this is what a model does? Machine learning fairness as a field is focused on the just treatment of individuals and groups under information processing with machine learning methods. While considerable attention has been given to mitigating discrimination of protected groups, vulnerable groups have not received the same attention. Unlike protected groups, which can be regarded as always vulnerable, a vulnerable group may be vulnerable in one context but not in another. This raises new challenges on how and when to protect vulnerable individuals and groups under machine learning. Methods from explainable artificial intelligence (XAI), in contrast, do consider more contextual issues and are concerned with answering the question "why was this decision made?". Neither existing fairness nor existing explainability methods allow us to ascertain if a prediction model identifies vulnerability. We discuss this problem and propose approaches for analysing prediction models in this respect.
Patrik Hammersborg, Inga Strümke
Self-trained autonomous agents developed using machine learning are showing great promise in a variety of control settings, perhaps most remarkably in applications involving autonomous vehicles. The main challenge associated with self-learned agents in the form of deep neural networks, is their black-box nature: it is impossible for humans to interpret deep neural networks. Therefore, humans cannot directly interpret the actions of deep neural network based agents, or foresee their robustness in different scenarios. In this work, we demonstrate a method for probing which concepts self-learning agents internalise in the course of their training. For demonstration, we use a chess playing agent in a fast and light environment developed specifically to be suitable for research groups without access to enormous computational resources or machine learning models.
Kimji N. Pellano, Inga Strümke, Daniel Groos et al.
Early detection of Cerebral Palsy (CP) is crucial for effective intervention and monitoring. This paper tests the reliability and applicability of Explainable AI (XAI) methods using a deep learning method that predicts CP by analyzing skeletal data extracted from video recordings of infant movements. Specifically, we use XAI evaluation metrics -- namely faithfulness and stability -- to quantitatively assess the reliability of Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM) in this specific medical application. We utilize a unique dataset of infant movements and apply skeleton data perturbations without distorting the original dynamics of the infant movements. Our CP prediction model utilizes an ensemble approach, so we evaluate the XAI metrics performances for both the overall ensemble and the individual models. Our findings indicate that both XAI methods effectively identify key body points influencing CP predictions and that the explanations are robust against minor data perturbations. Grad-CAM significantly outperforms CAM in the RISv metric, which measures stability in terms of velocity. In contrast, CAM performs better in the RISb metric, which relates to bone stability, and the RRS metric, which assesses internal representation robustness. Individual models within the ensemble show varied results, and neither CAM nor Grad-CAM consistently outperform the other, with the ensemble approach providing a representation of outcomes from its constituent models.
Patrik Hammersborg, Inga Strümke
Neural network models are widely used in a variety of domains, often as black-box solutions, since they are not directly interpretable for humans. The field of explainable artificial intelligence aims at developing explanation methods to address this challenge, and several approaches have been developed over the recent years, including methods for investigating what type of knowledge these models internalise during the training process. Among these, the method of concept detection, investigates which \emph{concepts} neural network models learn to represent in order to complete their tasks. In this work, we present an extension to the method of concept detection, named \emph{concept backpropagation}, which provides a way of analysing how the information representing a given concept is internalised in a given neural network model. In this approach, the model input is perturbed in a manner guided by a trained concept probe for the described model, such that the concept of interest is maximised. This allows for the visualisation of the detected concept directly in the input space of the model, which in turn makes it possible to see what information the model depends on for representing the described concept. We present results for this method applied to a various set of input modalities, and discuss how our proposed method can be used to visualise what information trained concept probes use, and the degree as to which the representation of the probed concept is entangled within the neural network model itself.
Felix Tempel, Espen Alexander F. Ihlen, Inga Strümke
The neurological condition known as cerebral palsy (CP) first manifests in infancy or early childhood and has a lifelong impact on motor coordination and body movement. CP is one of the leading causes of childhood disabilities, and early detection is crucial for providing appropriate treatment. However, such detection relies on assessments by human experts trained in methods like general movement assessment (GMA). These are not widely accessible, especially in developing countries. Conventional machine learning approaches offer limited predictive performance on CP detection tasks, and the approaches developed by the few available domain experts are generally dataset-specific, restricting their applicability beyond the context for which these were created. To address these challenges, we propose a neural architecture search (NAS) algorithm applying a reinforcement learning update scheme capable of efficiently optimizing for the best architectural and hyperparameter combination to discover the most suitable neural network configuration for detecting CP. Our method performs better on a real-world CP dataset than other approaches in the field, which rely on large ensembles. As our approach is less resource-demanding and performs better, it is particularly suitable for implementation in resource-constrained settings, including rural or developing areas with limited access to medical experts and the required diagnostic tools. The resulting model's lightweight architecture and efficient computation time allow for deployment on devices with limited processing power, reducing the need for expensive infrastructure, and can, therefore, be integrated into clinical workflows to provide timely and accurate support for early CP diagnosis.
Marte Eggen, Eirik Reiestad, Kristian Gjøsteen et al.
Recent cryptographic results establish that neural networks can be backdoored such that no efficient algorithm can distinguish them from a clean model. These guarantees, however, have been confined to stylised architectures of limited practical relevance, leaving open whether comparable undetectability extends to modern, end-to-end trained networks. We construct such an attack mechanism for state-of-the-art architectures, closely aligned to the cryptographic notion of undetectability, by identifying backdoor channels as learned latent directions, and show that the question of undetectability reduces to a hypothesis test between two unknown distributions over model parameters, which we conjecture to be intractable in practice. The consequence of this reframing is significant: if exploitable channels within a network's latent space are statistically indistinguishable from naturally learned directions, an attacker need not introduce foreign structure but can instead exploit the geometry the network already possesses. Demonstrating the approach on ResNet and Vision Transformer architectures trained on standard image classification datasets, the attack achieves both consistently high success rates with negligible clean accuracy degradation, and resists a comprehensive suite of post-training defences, none of which neutralise the backdoor without rendering the model unusable. Our results establish that cryptographic backdoors need not be artefacts requiring exotic architectures or artificial constructions, but identifiable as latent properties inherent to the geometry of learned representations.
Anna Rodum Bjøru, Jacob Lysnæs-Larsen, Oskar Jørgensen et al.
This work presents a conceptual framework for causal concept-based post-hoc Explainable Artificial Intelligence (XAI), based on the requirements that explanations for non-interpretable models should be understandable as well as faithful to the model being explained. Local and global explanations are generated by calculating the probability of sufficiency of concept interventions. Example explanations are presented, generated with a proof-of-concept model made to explain classifiers trained on the CelebA dataset. Understandability is demonstrated through a clear concept-based vocabulary, subject to an implicit causal interpretation. Fidelity is addressed by highlighting important framework assumptions, stressing that the context of explanation interpretation must align with the context of explanation generation.
Kimji N. Pellano, Inga Strümke, Espen Alexander F. Ihlen
The advancement of deep learning in human activity recognition (HAR) using 3D skeleton data is critical for applications in healthcare, security, sports, and human-computer interaction. This paper tackles a well-known gap in the field, which is the lack of testing in the applicability and reliability of XAI evaluation metrics in the skeleton-based HAR domain. We have tested established XAI metrics namely faithfulness and stability on Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM) to address this problem. The study also introduces a perturbation method that respects human biomechanical constraints to ensure realistic variations in human movement. Our findings indicate that \textit{faithfulness} may not be a reliable metric in certain contexts, such as with the EfficientGCN model. Conversely, stability emerges as a more dependable metric when there is slight input data perturbations. CAM and Grad-CAM are also found to produce almost identical explanations, leading to very similar XAI metric performance. This calls for the need for more diversified metrics and new XAI methods applied in skeleton-based HAR.
Felix Tempel, Espen Alexander F. Ihlen, Lars Adde et al.
In Human Activity Recognition (HAR), understanding the intricacy of body movements within high-risk applications is essential. This study uses SHapley Additive exPlanations (SHAP) to explain the decision-making process of Graph Convolution Networks (GCNs) when classifying activities with skeleton data. We employ SHAP to explain two real-world datasets: one for cerebral palsy (CP) classification and the widely used NTU RGB+D 60 action recognition dataset. To test the explanation, we introduce a novel perturbation approach that modifies the model's edge importance matrix, allowing us to evaluate the impact of specific body key points on prediction outcomes. To assess the fidelity of our explanations, we employ informed perturbation, targeting body key points identified as important by SHAP and comparing them against random perturbation as a control condition. This perturbation enables a judgment on whether the body key points are truly influential or non-influential based on the SHAP values. Results on both datasets show that body key points identified as important through SHAP have the largest influence on the accuracy, specificity, and sensitivity metrics. Our findings highlight that SHAP can provide granular insights into the input feature contribution to the prediction outcome of GCNs in HAR tasks. This demonstrates the potential for more interpretable and trustworthy models in high-stakes applications like healthcare or rehabilitation.
Luis M. Lopez-Ramos, Florian Leiser, Aditya Rastogi et al.
The joint implementation of federated learning (FL) and explainable artificial intelligence (XAI) could allow training models from distributed data and explaining their inner workings while preserving essential aspects of privacy. Toward establishing the benefits and tensions associated with their interplay, this scoping review maps the publications that jointly deal with FL and XAI, focusing on publications that reported an interplay between FL and model interpretability or post-hoc explanations. Out of the 37 studies meeting our criteria, only one explicitly and quantitatively analyzed the influence of FL on model explanations, revealing a significant research gap. The aggregation of interpretability metrics across FL nodes created generalized global insights at the expense of node-specific patterns being diluted. Several studies proposed FL algorithms incorporating explanation methods to safeguard the learning process against defaulting or malicious nodes. Studies using established FL libraries or following reporting guidelines are a minority. More quantitative research and structured, transparent practices are needed to fully understand their mutual impact and under which conditions it happens.
Felix Tempel, Daniel Groos, Espen Alexander F. Ihlen et al.
Explaining machine learning (ML) models using eXplainable AI (XAI) techniques has become essential to make them more transparent and trustworthy. This is especially important in high-stakes domains like healthcare, where understanding model decisions is critical to ensure ethical, sound, and trustworthy outcome predictions. However, users are often confused about which explanability method to choose for their specific use case. We present a comparative analysis of widely used explainability methods, Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), within the domain of human activity recognition (HAR) utilizing graph convolutional networks (GCNs). By evaluating these methods on skeleton-based data from two real-world datasets, including a healthcare-critical cerebral palsy (CP) case, this study provides vital insights into both approaches' strengths, limitations, and differences, offering a roadmap for selecting the most appropriate explanation method based on specific models and applications. We quantitatively and quantitatively compare these methods, focusing on feature importance ranking, interpretability, and model sensitivity through perturbation experiments. While SHAP provides detailed input feature attribution, Grad-CAM delivers faster, spatially oriented explanations, making both methods complementary depending on the application's requirements. Given the importance of XAI in enhancing trust and transparency in ML models, particularly in sensitive environments like healthcare, our research demonstrates how SHAP and Grad-CAM could complement each other to provide more interpretable and actionable model explanations.
Felix Tempel, Inga Strümke, Espen Alexander F. Ihlen
This paper introduces AutoGCN, a generic Neural Architecture Search (NAS) algorithm for Human Activity Recognition (HAR) using Graph Convolution Networks (GCNs). HAR has gained attention due to advances in deep learning, increased data availability, and enhanced computational capabilities. At the same time, GCNs have shown promising results in modeling relationships between body key points in a skeletal graph. While domain experts often craft dataset-specific GCN-based methods, their applicability beyond this specific context is severely limited. AutoGCN seeks to address this limitation by simultaneously searching for the ideal hyperparameters and architecture combination within a versatile search space using a reinforcement controller while balancing optimal exploration and exploitation behavior with a knowledge reservoir during the search process. We conduct extensive experiments on two large-scale datasets focused on skeleton-based action recognition to assess the proposed algorithm's performance. Our experimental results underscore the effectiveness of AutoGCN in constructing optimal GCN architectures for HAR, outperforming conventional NAS and GCN methods, as well as random search. These findings highlight the significance of a diverse search space and an expressive input representation to enhance the network performance and generalizability.
Inga Strümke, Helge Langseth
Diffusion models are loosely modelled based on non-equilibrium thermodynamics, where \textit{diffusion} refers to particles flowing from high-concentration regions towards low-concentration regions. In statistics, the meaning is quite similar, namely the process of transforming a complex distribution $p_{\text{complex}}$ on $\mathbb{R}^d$ to a simple distribution $p_{\text{prior}}$ on the same domain. This constitutes a Markov chain of diffusion steps of slowly adding random noise to data, followed by a reverse diffusion process in which the data is reconstructed from the noise. The diffusion model learns the data manifold to which the original and thus the reconstructed data samples belong, by training on a large number of data points. While the diffusion process pushes a data sample off the data manifold, the reverse process finds a trajectory back to the data manifold. Diffusion models have -- unlike variational autoencoder and flow models -- latent variables with the same dimensionality as the original data, and they are currently\footnote{At the time of writing, 2023.} outperforming other approaches -- including Generative Adversarial Networks (GANs) -- to modelling the distribution of, e.g., natural images.
Jacob Lysnæs-Larsen, Marte Eggen, Inga Strümke
In explainable AI, Concept Activation Vectors (CAVs) are typically obtained by training linear classifier probes to detect human-understandable concepts as directions in the activation space of deep neural networks. It is widely assumed that a high probe accuracy indicates a CAV faithfully representing its target concept. However, we show that the probe's classification accuracy alone is an unreliable measure of concept alignment, i.e., the degree to which a CAV captures the intended concept. In fact, we argue that probes are more likely to capture spurious correlations than they are to represent only the intended concept. As part of our analysis, we demonstrate that deliberately misaligned probes constructed to exploit spurious correlations, achieve an accuracy close to that of standard probes. To address this severe problem, we introduce a novel concept localization method based on spatial linear attribution, and provide a comprehensive comparison of it to existing feature visualization techniques for detecting and mitigating concept misalignment. We further propose three classes of metrics for quantitatively assessing concept alignment: hard accuracy, segmentation scores, and augmentation robustness. Our analysis shows that probes with translation invariance and spatial alignment consistently increase concept alignment. These findings highlight the need for alignment-based evaluation metrics rather than probe accuracy, and the importance of tailoring probes to both the model architecture and the nature of the target concept.
Marte Eggen, Jacob Lysnæs-Larsen, Inga Strümke
The attention mechanism lies at the core of the transformer architecture, providing an interpretable model-internal signal that has motivated a growing interest in attention-based model explanations. Although attention weights do not directly determine model outputs, they reflect patterns of token influence that can inform and complement established explainability techniques. This work studies the potential of utilising the information encoded in attention weights to provide meaningful model explanations by integrating them into explainable AI (XAI) frameworks that target fundamentally different aspects of model behaviour. To this end, we develop two novel explanation methods applicable to both natural language processing and computer vision tasks. The first integrates attention weights into the Shapley value decomposition by redefining the characteristic function in terms of pairwise token interactions via attention weights, thus adapting this widely used game-theoretic solution concept to provide attention-driven attributions for local explanations. The second incorporates attention weights into token-level directional derivatives defined through concept activation vectors to measure concept sensitivity for global explanations. Our empirical evaluations on standard benchmarks and in a comparison study with widely used explanation methods show that attention weights can be meaningfully incorporated into the studied XAI frameworks, highlighting their value in enriching transformer explainability.
Kimji N. Pellano, Inga Strümke, Daniel Groos et al.
Cerebral Palsy (CP) is a prevalent motor disability in children, for which early detection can significantly improve treatment outcomes. While skeleton-based Graph Convolutional Network (GCN) models have shown promise in automatically predicting CP risk from infant videos, their "black-box" nature raises concerns about clinical explainability. To address this, we introduce a perturbation framework tailored for infant movement features and use it to compare two explainable AI (XAI) methods: Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM). First, we identify significant and non-significant body keypoints in very low- and very high-risk infant video snippets based on the XAI attribution scores. We then conduct targeted velocity and angular perturbations, both individually and in combination, on these keypoints to assess how the GCN model's risk predictions change. Our results indicate that velocity-driven features of the arms, hips, and legs have a dominant influence on CP risk predictions, while angular perturbations have a more modest impact. Furthermore, CAM and Grad-CAM show partial convergence in their explanations for both low- and high-risk CP groups. Our findings demonstrate the use of XAI-driven movement analysis for early CP prediction and offer insights into potential movement-based biomarker discovery that warrant further clinical validation.
Tannista Banerjee, Ayan Paul, Vishak Srikanth et al.
The analysis of causation is a challenging task that can be approached in various ways. With the increasing use of machine learning based models in computational socioeconomics, explaining these models while taking causal connections into account is a necessity. In this work, we advocate the use of an explanatory framework from cooperative game theory augmented with $do$ calculus, namely causal Shapley values. Using causal Shapley values, we analyze socioeconomic disparities that have a causal link to the spread of COVID-19 in the USA. We study several phases of the disease spread to show how the causal connections change over time. We perform a causal analysis using random effects models and discuss the correspondence between the two methods to verify our results. We show the distinct advantages a non-linear machine learning models have over linear models when performing a multivariate analysis, especially since the machine learning models can map out non-linear correlations in the data. In addition, the causal Shapley values allow for including the causal structure in the variable importance computed for the machine learning model.
Sindre Benjamin Remman, Inga Strümke, Anastasios M. Lekkas
We investigate the effect of including domain knowledge about a robotic system's causal relations when generating explanations. To this end, we compare two methods from explainable artificial intelligence, the popular KernelSHAP and the recent causal SHAP, on a deep neural network trained using deep reinforcement learning on the task of controlling a lever using a robotic manipulator. A primary disadvantage of KernelSHAP is that its explanations represent only the features' direct effects on a model's output, not considering the indirect effects a feature can have on the output by affecting other features. Causal SHAP uses a partial causal ordering to alter KernelSHAP's sampling procedure to incorporate these indirect effects. This partial causal ordering defines the causal relations between the features, and we specify this using domain knowledge about the lever control task. We show that enabling an explanation method to account for indirect effects and incorporating some domain knowledge can lead to explanations that better agree with human intuition. This is especially favorable for a real-world robotics task, where there is considerable causality at play, and in addition, the required domain knowledge is often handily available.
Pål Vegard Johnsen, Inga Strümke, Signe Riemer-Sørensen et al.
Estimating feature importance is a significant aspect of explaining data-based models. Besides explaining the model itself, an equally relevant question is which features are important in the underlying data generating process. We present a Shapley value based framework for inferring the importance of individual features, including uncertainty in the estimator. We build upon the recently published feature importance measure of SAGE (Shapley additive global importance) and introduce sub-SAGE which can be estimated without resampling for tree-based models. We argue that the uncertainties can be estimated from bootstrapping and demonstrate the approach for tree ensemble methods. The framework is exemplified on synthetic data as well as high-dimensional genomics data.
Andrea M. Storås, Inga Strümke, Michael A. Riegler et al.
Dry eye disease (DED) has a prevalence of between 5 and 50\%, depending on the diagnostic criteria used and population under study. However, it remains one of the most underdiagnosed and undertreated conditions in ophthalmology. Many tests used in the diagnosis of DED rely on an experienced observer for image interpretation, which may be considered subjective and result in variation in diagnosis. Since artificial intelligence (AI) systems are capable of advanced problem solving, use of such techniques could lead to more objective diagnosis. Although the term `AI' is commonly used, recent success in its applications to medicine is mainly due to advancements in the sub-field of machine learning, which has been used to automatically classify images and predict medical outcomes. Powerful machine learning techniques have been harnessed to understand nuances in patient data and medical images, aiming for consistent diagnosis and stratification of disease severity. This is the first literature review on the use of AI in DED. We provide a brief introduction to AI, report its current use in DED research and its potential for application in the clinic. Our review found that AI has been employed in a wide range of DED clinical tests and research applications, primarily for interpretation of interferometry, slit-lamp and meibography images. While initial results are promising, much work is still needed on model development, clinical testing and standardisation.
Daniel Alvestad, Nikolai Fomin, Jörn Kersten et al.
We investigate enhancing the sensitivity of new physics searches at the LHC by machine learning in the case of background dominance and a high degree of overlap between the observables for signal and background. We use two different models, XGBoost and a deep neural network, to exploit correlations between observables and compare this approach to the traditional cut-and-count method. We consider different methods to analyze the models' output, finding that a template fit generally performs better than a simple cut. By means of a Shapley decomposition, we gain additional insight into the relationship between event kinematics and the machine learning model output. We consider a supersymmetric scenario with a metastable sneutrino as a concrete example, but the methodology can be applied to a much wider class of models.
Inga Strümke, Marija Slavkovik, Vince I. Madai
While the demand for ethical artificial intelligence (AI) systems increases, the number of unethical uses of AI accelerates, even though there is no shortage of ethical guidelines. We argue that a possible underlying cause for this is that AI developers face a social dilemma in AI development ethics, preventing the widespread adaptation of ethical best practices. We define the social dilemma for AI development and describe why the current crisis in AI development ethics cannot be solved without relieving AI developers of their social dilemma. We argue that AI development must be professionalised to overcome the social dilemma, and discuss how medicine can be used as a template in this process.
Daniel Fryer, Inga Strümke, Hien Nguyen
The Shapley value has become popular in the Explainable AI (XAI) literature, thanks, to a large extent, to a solid theoretical foundation, including four "favourable and fair" axioms for attribution in transferable utility games. The Shapley value is provably the only solution concept satisfying these axioms. In this paper, we introduce the Shapley value and draw attention to its recent uses as a feature selection tool. We call into question this use of the Shapley value, using simple, abstract "toy" counterexamples to illustrate that the axioms may work against the goals of feature selection. From this, we develop a number of insights that are then investigated in concrete simulation settings, with a variety of Shapley value formulations, including SHapley Additive exPlanations (SHAP) and Shapley Additive Global importancE (SAGE).
Daniel Vidali Fryer, Inga Strümke, Hien Nguyen
Shapley values have become increasingly popular in the machine learning literature thanks to their attractive axiomatisation, flexibility, and uniqueness in satisfying certain notions of `fairness'. The flexibility arises from the myriad potential forms of the Shapley value \textit{game formulation}. Amongst the consequences of this flexibility is that there are now many types of Shapley values being discussed, with such variety being a source of potential misunderstanding. To the best of our knowledge, all existing game formulations in the machine learning and statistics literature fall into a category which we name the model-dependent category of game formulations. In this work, we consider an alternative and novel formulation which leads to the first instance of what we call model-independent Shapley values. These Shapley values use a (non-parametric) measure of non-linear dependence as the characteristic function. The strength of these Shapley values is in their ability to uncover and attribute non-linear dependencies amongst features. We introduce and demonstrate the use of the energy distance correlations, affine-invariant distance correlation, and Hilbert-Shmidt independence criterion as Shapley value characteristic functions. In particular, we demonstrate their potential value for exploratory data analysis and model diagnostics. We conclude with an interesting expository application to a classical medical survey data set.