The use of machine learning (ML) in critical domains such as medicine poses risks and requires regulation. One requirement is that decisions of ML systems in high-risk applications should be human-understandable. The field of "explainable artificial intelligence" (XAI) seemingly addresses this need. However, in its current form, XAI is unfit to provide quality control for ML; it itself needs scrutiny. Popular XAI methods cannot reliably answer important questions about ML models, their training data, or a given test input. We recapitulate results demonstrating that popular XAI methods systematically attribute importance to input features that are independent of the prediction target. This limits their utility for purposes such as model and data (in)validation, model improvement, and scientific discovery. We argue that the fundamental reason for this limitation is that current XAI methods do not address well-defined problems and are not evaluated against objective criteria of explanation correctness. Researchers should formally define the problems they intend to solve first and then design methods accordingly. This will lead to notions of explanation correctness that can be theoretically verified and objective metrics of explanation performance that can be assessed using ground-truth data.

Large pre-trained language models have become popular for many applications and form an important backbone of many downstream tasks in natural language processing (NLP). Applying 'explainable artificial intelligence' (XAI) techniques to enrich such models' outputs is considered crucial for assuring their quality and shedding light on their inner workings. However, large language models are trained on a plethora of data containing a variety of biases, such as gender biases, affecting model weights and, potentially, behavior. Currently, it is unclear to what extent such biases also impact model explanations in possibly unfavorable ways. We create a gender-controlled text dataset, GECO, in which otherwise identical sentences appear in male and female forms. This gives rise to ground-truth 'world explanations' for gender classification tasks, enabling the objective evaluation of the correctness of XAI methods. We also provide GECOBench, a rigorous quantitative evaluation framework benchmarking popular XAI methods, applying them to pre-trained language models fine-tuned to different degrees. This allows us to investigate how pre-training induces undesirable bias in model explanations and to what extent fine-tuning can mitigate such explanation bias. We show a clear dependency between explanation performance and the number of fine-tuned layers, where XAI methods are observed to particularly benefit from fine-tuning or complete retraining of embedding layers. Remarkably, this relationship holds for models achieving similar classification performance on the same task. With that, we highlight the utility of the proposed gender-controlled dataset and novel benchmarking approach for research and development of novel XAI methods. All code including dataset generation, model training, evaluation and visualization is available at: https://github.com/braindatalab/gecobench

The evolving landscape of explainable artificial intelligence (XAI) aims to improve the interpretability of intricate machine learning (ML) models, yet faces challenges in formalisation and empirical validation, being an inherently unsupervised process. In this paper, we bring together various benchmark datasets and novel performance metrics in an initial benchmarking platform, the Explainable AI Comparison Toolkit (EXACT), providing a standardised foundation for evaluating XAI methods. Our datasets incorporate ground truth explanations for class-conditional features, and leveraging novel quantitative metrics, this platform assesses the performance of post-hoc XAI methods in the quality of the explanations they produce. Our recent findings have highlighted the limitations of popular XAI methods, as they often struggle to surpass random baselines, attributing significance to irrelevant features. Moreover, we show the variability in explanations derived from different equally performing model architectures. This initial benchmarking platform therefore aims to allow XAI researchers to test and assure the high quality of their newly developed methods.

Only recently, a trend towards the objective empirical validation of XAI methods using ground truth data has been observed (Tjoa and Guan, 2020; Li et al, 2021; Zhou et al, 2022; Arras et al, 2022; Gevaert et al, 2022; Agarwal et al, 2022). These studies are, however, limited in the extent to which they permit a quantitative assessment of explanation performance, in the breadth of XAI methods evaluated, and in the difficulty of the posed'explanation' problems. In particular, most published benchmark datasets are constructed in a way such that realistic correlations between class-dependent (e.g., the foreground or object of an image) and class-agnostic (e.g., the image background) features are excluded. In practice, such dependencies can give rise to features acting as suppressor variables. Briefly, suppressor variables have no statistical association to the prediction target on their own, yet including them may allow an ML model to remove unwanted signals (noise), which can lead to improved predictions. In the context of image or photography data, suppressor variables could be parts of the background that capture the general lighting conditions.

Convolutional Neural Networks (CNNs) are frequently and successfully used in medical prediction tasks. They are often used in combination with transfer learning, leading to improved performance when training data for the task are scarce. The resulting models are highly complex and typically do not provide any insight into their predictive mechanisms, motivating the field of 'explainable' artificial intelligence (XAI). However, previous studies have rarely quantitatively evaluated the 'explanation performance' of XAI methods against ground-truth data, and transfer learning and its influence on objective measures of explanation performance has not been investigated. Here, we propose a benchmark dataset that allows for quantifying explanation performance in a realistic magnetic resonance imaging (MRI) classification task. We employ this benchmark to understand the influence of transfer learning on the quality of explanations. Experimental results show that popular XAI methods applied to the same underlying model differ vastly in performance, even when considering only correctly classified examples. We further observe that explanation performance strongly depends on the task used for pre-training and the number of CNN layers pre-trained. These results hold after correcting for a substantial correlation between explanation and classification performance.

In recent years, the community of 'explainable artificial intelligence' (XAI) has created a vast body of methods to bridge a perceived gap between model 'complexity' and 'interpretability'. However, a concrete problem to be solved by XAI methods has not yet been formally stated. As a result, XAI methods are lacking theoretical and empirical evidence for the 'correctness' of their explanations, limiting their potential use for quality-control and transparency purposes. At the same time, Haufe et al. (2014) showed, using simple toy examples, that even standard interpretations of linear models can be highly misleading. Specifically, high importance may be attributed to so-called suppressor variables lacking any statistical relation to the prediction target. This behavior has been confirmed empirically for a large array of XAI methods in Wilming et al. (2022). Here, we go one step further by deriving analytical expressions for the behavior of a variety of popular XAI methods on a simple two-dimensional binary classification problem involving Gaussian class-conditional distributions. We show that the majority of the studied approaches will attribute non-zero importance to a non-class-related suppressor feature in the presence of correlated noise. This poses important limitations on the interpretations and conclusions that the outputs of these XAI methods can afford.

Machine learning (ML) is increasingly often used to inform high-stakes decisions. As complex ML models (e.g., deep neural networks) are often considered black boxes, a wealth of procedures has been developed to shed light on their inner workings and the ways in which their predictions come about, defining the field of 'explainable AI' (XAI). Saliency methods rank input features according to some measure of 'importance'. Such methods are difficult to validate since a formal definition of feature importance is, thus far, lacking. It has been demonstrated that some saliency methods can highlight features that have no statistical association with the prediction target (suppressor variables). To avoid misinterpretations due to such behavior, we propose the actual presence of such an association as a necessary condition and objective preliminary definition for feature importance. We carefully crafted a ground-truth dataset in which all statistical dependencies are well-defined and linear, serving as a benchmark to study the problem of suppressor variables. We evaluate common explanation methods including LRP, DTD, PatternNet, PatternAttribution, LIME, Anchors, SHAP, and permutation-based methods with respect to our objective definition. We show that most of these methods are unable to distinguish important features from suppressors in this setting.