Causal discovery, or identifying causal relationships from observational data, is a notoriously challenging task, with numerous methods proposed to tackle it. Despite this, in-the-wild evaluation of these methods is still lacking, as works frequently rely on synthetic data evaluation and sparse real-world examples under critical theoretical assumptions. Real-world causal structures, however, are often complex, evolving over time, non-linear, and influenced by unobserved factors, making it hard to decide on a proper causal discovery strategy. To bridge this gap, we introduce CausalRivers, the largest in-the-wild causal discovery benchmarking kit for time-series data to date. CausalRivers features an extensive dataset on river discharge that covers the eastern German territory (666 measurement stations) and the state of Bavaria (494 measurement stations). It spans the years 2019 to 2023 with a 15-minute temporal resolution. Further, we provide additional data from a flood around the Elbe River, as an event with a pronounced distributional shift. Leveraging multiple sources of information and time-series meta-data, we constructed two distinct causal ground truth graphs (Bavaria and eastern Germany). These graphs can be sampled to generate thousands of subgraphs to benchmark causal discovery across diverse and challenging settings. To demonstrate the utility of CausalRivers, we evaluate several causal discovery approaches through a set of experiments to identify areas for improvement. CausalRivers has the potential to facilitate robust evaluations and comparisons of causal discovery methods. Besides this primary purpose, we also expect that this dataset will be relevant for connected areas of research, such as time-series forecasting and anomaly detection. Based on this, we hope to push benchmark-driven method development that fosters advanced techniques for causal discovery, as is the case for many other areas of machine learning.

Concepts such as objects, patterns, and shapes are how humans understand the world. Building on this intuition, concept-based explainability methods aim to study representations learned by deep neural networks in relation to human-understandable concepts. Here, Concept Activation Vectors (CAVs) are an important tool and can identify whether a model learned a concept or not. However, the computational cost and time requirements of existing CAV computation pose a significant challenge, particularly in large-scale, high-dimensional architectures. To address this limitation, we introduce \methodname, a novel approach that accelerates the extraction of CAVs by up to \maxspeedup (on average \avgspeedup). %times. We provide a theoretical foundation for our approach and give concrete assumptions under which it is equivalent to established SVM-based methods. Our empirical results demonstrate that CAVs calculated with \methodname maintain similar performance while being more efficient and stable. In downstream applications, i.e., concept-based explanation methods, we show that \methodname can act as a replacement leading to equivalent insights. Hence, our approach enables previously infeasible investigations of deep models, which we demonstrate by tracking the evolution of concepts during model training.

Due to the data-centric approach of modern machine learning, biases present in the training data are frequently learned by deep models. It is often necessary to collect new data and retrain the models from scratch to remedy these issues, which can be expensive in critical areas such as medicine. We investigate whether it is possible to fix pre-trained model behavior using very few unbiased examples. We show that we can improve performance by tuning the models while penalizing parameter changes. Hence, we are keeping pre-trained knowledge while simultaneously correcting the harmful behavior. Toward this goal, we tune a zero-initialized copy of the frozen pre-trained network using strong parameter norms. Secondly, we introduce an early stopping scheme to modify baselines and reduce overfitting. Our approaches lead to improvements in four datasets common in the debiasing and domain shift literature. We especially see benefits in an iterative setting, where new samples are added continuously. Hence, we demonstrate the effectiveness of tuning while penalizing change to fix pre-trained models without retraining from scratch.

Deep learning models achieve high predictive performance but lack intrinsic interpretability, hindering our understanding of the learned prediction behavior. Existing local explainability methods focus on associations, neglecting the causal drivers of model predictions. Other approaches adopt a causal perspective but primarily provide more general global explanations. However, for specific inputs, it's unclear whether globally identified factors apply locally. To address this limitation, we introduce a novel framework for local interventional explanations by leveraging recent advances in image-to-image editing models. Our approach performs gradual interventions on semantic properties to quantify the corresponding impact on a model's predictions using a novel score, the expected property gradient magnitude. We demonstrate the effectiveness of our approach through an extensive empirical evaluation on a wide range of architectures and tasks. First, we validate it in a synthetic scenario and demonstrate its ability to locally identify biases. Afterward, we apply our approach to analyze network training dynamics, investigate medical skin lesion classifiers, and study a pre-trained CLIP model with real-life interventional data. Our results highlight the potential of interventional explanations on the property level to reveal new insights into the behavior of deep models.

Models like DALL-E and Stable Diffusion substantially improved image generation. While visually convincing, they still exhibit distinct differences compared to real-world images and desired image distributions. Previous studies already identified issues at a spectral level. We also find discrepancies in terms of style authenticity and aesthetics. Based on these insights we investigate three distinct strategies to modify such models and enhance their alignment for selected image distributions of artworks. First, prompt optimization can be done without updating the model, but has limited efficacy. Fine-tuning the U-Net component can enhance denoising capabilities, leading to improved image quality. Finally, modifying the image decoder can help correct spectral misalignments. Our experiments on the ArtEmis dataset using three complementary measures show that high-frequency artifacts remain challenging, but alignment can usually be improved by at least 10%.

A considerable amount of research has been dedicated to creating systems that aid medical professionals in labor-intensive early screening tasks, which, to this date, often leverage convolutional deep-learning architectures. Recently, several studies have explored the application of self-attention mechanisms in the field of computer vision. These studies frequently demonstrate empirical improvements over traditional, fully convolutional approaches across a range of datasets and tasks. To assess this trend for medical imaging, we enhance two commonly used convolutional architectures with various self-attention mechanisms and evaluate them on two distinct medical datasets. We compare these enhanced architectures with similarly sized convolutional and attention-based baselines and rigorously assess performance gains through statistical evaluation. Furthermore, we investigate how the inclusion of self-attention influences the features learned by these models by assessing global and local explanations of model behavior. Contrary to our expectations, after performing an appropriate hyperparameter search, self-attention-enhanced architectures show no significant improvements in balanced accuracy compared to the evaluated baselines. Further, we find that relevant global features like dermoscopic structures in skin lesion images are not properly learned by any architecture. Finally, by assessing local explanations, we find that the inherent interpretability of self-attention mechanisms does not provide additional insights. Out-of-the-box model-agnostic approaches can provide explanations that are similar or even more faithful to the actual model behavior. We conclude that simply integrating attention mechanisms is unlikely to lead to a consistent increase in performance compared to fully convolutional methods in medical imaging applications.

Deep models trained on large amounts of data often incorporate implicit biases present during training time. If later such a bias is discovered during inference or deployment, it is often necessary to acquire new data and retrain the model. This behavior is especially problematic in critical areas such as autonomous driving or medical decision-making. In these scenarios, new data is often expensive and hard to come by. In this work, we present a method based on change penalization that takes a pre-trained model and adapts the weights to mitigate a previously detected bias. We achieve this by tuning a zero-initialized copy of a frozen pre-trained network. Our method needs very few, in extreme cases only a single, examples that contradict the bias to increase performance. Additionally, we propose an early stopping criterion to modify baselines and reduce overfitting. We evaluate our approach on a well-known bias in skin lesion classification and three other datasets from the domain shift literature. We find that our approach works especially well with very few images. Simple fine-tuning combined with our early stopping also leads to performance benefits for a larger number of tuning samples.

Understanding expressions is vital for deciphering human behavior, and nowadays, end-to-end trained black box models achieve high performance. Due to the black-box nature of these models, it is unclear how they behave when applied out-of-distribution. Specifically, these models show decreased performance for unilateral facial palsy patients. We hypothesize that one crucial factor guiding the internal decision rules is facial symmetry. In this work, we use insights from causal reasoning to investigate the hypothesis. After deriving a structural causal model, we develop a synthetic interventional framework. This approach allows us to analyze how facial symmetry impacts a network's output behavior while keeping other factors fixed. All 17 investigated expression classifiers significantly lower their output activations for reduced symmetry. This result is congruent with observed behavior on real-world data from healthy subjects and facial palsy patients. As such, our investigation serves as a case study for identifying causal factors that influence the behavior of black-box models.

Facial expression-based human emotion recognition is a critical research area in psychology and medicine. State-of-the-art classification performance is only reached by end-to-end trained neural networks. Nevertheless, such black-box models lack transparency in their decisionmaking processes, prompting efforts to ascertain the rules that underlie classifiers' decisions. Analyzing single inputs alone fails to expose systematic learned biases. These biases can be characterized as facial properties summarizing abstract information like age or medical conditions. Therefore, understanding a model's prediction behaviorrequires an analysis rooted in causality along such selected properties. We demonstrate that up to 91.25\% of classifier output behavior changes are statistically significant concerning basic properties. Among those are age, gender, and facial symmetry. Furthermore, the medical usage of surface electromyography significantly influences emotion prediction. We introduce a workflow to evaluate explicit properties and their impact. These insights might help medical professionals select and apply classifiers regarding their specialized data and properties.

A substantial body of research has focused on developing systems that assist medical professionals during labor-intensive early screening processes, many based on convolutional deep-learning architectures. Recently, multiple studies explored the application of so-called self-attention mechanisms in the vision domain. These studies often report empirical improvements over fully convolutional approaches on various datasets and tasks. To evaluate this trend for medical imaging, we extend two widely adopted convolutional architectures with different self-attention variants on two different medical datasets. With this, we aim to specifically evaluate the possible advantages of additional self-attention. We compare our models with similarly sized convolutional and attention-based baselines and evaluate performance gains statistically. Additionally, we investigate how including such layers changes the features learned by these models during the training. Following a hyperparameter search, and contrary to our expectations, we observe no significant improvement in balanced accuracy over fully convolutional models. We also find that important features, such as dermoscopic structures in skin lesion images, are still not learned by employing self-attention. Finally, analyzing local explanations, we confirm biased feature usage. We conclude that merely incorporating attention is insufficient to surpass the performance of existing fully convolutional methods.

It is common for domain experts like physicians in medical studies to examine features for their reliability with respect to a specific domain task. When introducing machine learning, a common expectation is that machine learning models use the same features as human experts to solve a task but that is not always the case. Moreover, datasets often contain features that are known from domain knowledge to generalize badly to the real world, referred to as biases. Current debiasing methods only remove such influences. To additionally integrate the domain knowledge about well-established features into the training of a model, their relevance should be increased. We present a method that permits the manipulation of the relevance of features by actively steering the model's feature selection during the training process. That is, it allows both the discouragement of biases and encouragement of well-established features to incorporate domain knowledge about the feature reliability. We model our objectives for actively steering the feature selection process as a constrained optimization problem, which we implement via a loss regularization that is based on batch-wise feature attributions. We evaluate our approach on a novel synthetic regression dataset and a computer vision dataset. We observe that it successfully steers the features a model selects during the training process. This is a strong indicator that our method can be used to integrate domain knowledge about well-established features into a model.

The performance of a machine learning model is characterized by its ability to accurately represent the input-output relationship and its behavior on unseen data. A prerequisite for high performance is that causal relationships of features with the model outcome are correctly represented. This work analyses the causal relationships by investigating the relevance of features in machine learning models using conditional independence tests. For this, an attribution method based on Pearl's causality framework is employed.Our presented approach analyzes two data-driven models designed for the harvest-readiness prediction of cauliflower plants: one base model and one model where the decision process is adjusted based on local explanations. Additionally, we propose a visualization technique inspired by Partial Dependence Plots to gain further insights into the model behavior. The experiments presented in this paper find that both models learn task-relevant features during fine-tuning when compared to the ImageNet pre-trained weights. However, both models differ in their feature relevance, specifically in whether they utilize the image recording date. The experiments further show that our approach is able to reveal that the adjusted model is able to reduce the trends for the observed biases. Furthermore, the adjusted model maintains the desired behavior for the semantically meaningful feature of cauliflower head diameter, predicting higher harvest-readiness scores for higher feature realizations, which is consistent with existing domain knowledge. The proposed investigation approach can be applied to other domain-specific tasks to aid practitioners in evaluating model choices.

In many domains, so-called foundation models were recently proposed. These models are trained on immense amounts of data resulting in impressive performances on various downstream tasks and benchmarks. Later works focus on leveraging this pre-trained knowledge by combining these models. To reduce data and compute requirements, we utilize and combine foundation models in two ways. First, we use language and vision models to extract and generate a challenging language vision task in the form of artwork interpretation pairs. Second, we combine and fine-tune CLIP as well as GPT-2 to reduce compute requirements for training interpretation models. We perform a qualitative and quantitative analysis of our data and conclude that generating artwork leads to improvements in visual-text alignment and, therefore, to more proficient interpretation models. Our approach addresses how to leverage and combine pre-trained models to tackle tasks where existing data is scarce or difficult to obtain.

There are still open questions about how the learned representations of deep models change during the training process. Understanding this process could aid in validating the training. Towards this goal, previous works analyze the training in the mutual information plane. We use a different approach and base our analysis on a method built on Reichenbach’s common cause principle. Using this method, we test whether the model utilizes information contained in human-defined features. Given such a set of features, we investigate how the relative feature usage changes throughout the training process. We analyze mul- tiple networks training on different tasks, including melanoma classifica- tion as a real-world application. We find that over the training, models concentrate on features containing information relevant to the task. This concentration is a form of representation compression. Crucially, we also find that the selected features can differ between training from-scratch and finetuning a pre-trained network.

Skin cancer is the most common form of cancer, and melanoma is the leading cause of cancer related deaths. To improve the chances of survival, early detection of melanoma is crucial. Automated systems for classifying skin lesions can assist with initial analysis. However, if we expect people to entrust their well-being to an automatic classification algorithm, it is important to ensure that the algorithm makes medically sound decisions. We investigate this question by testing whether two state-of-the-art models use the features defined in the dermoscopic ABCD rule or whether they rely on biases. We use a method that frames supervised learning as a structural causal model, thus reducing the question whether a feature is used to a conditional dependence test. We show that this conditional dependence method yields meaningful results on data from the ISIC archive. Furthermore, we find that the selected models incorporate asymmetry, border and dermoscopic structures in their decisions but not color. Finally, we show that the same classifiers also use bias features such as the patient's age, skin color or the existence of colorful patches.

Uncertainty sampling is a widely used active learning strategy to select unlabeled examples for annotation. However, previous work hints at weaknesses of uncertainty sampling when combined with deep learning, where the amount of data is even more significant. To investigate these problems, we analyze the properties of the latent statistical estimators of uncertainty sampling in simple scenarios. We prove that uncertainty sampling converges towards some decision boundary. Additionally, we show that it can be inconsistent, leading to incorrect estimates of the optimal latent boundary. The inconsistency depends on the latent class distribution, more specifically on the class overlap. Further, we empirically analyze the variance of the decision boundary and find that the performance of uncertainty sampling is also connected to the class regions overlap. We argue that our findings could be the first step towards explaining the poor performance of uncertainty sampling combined with deep models.