Sebastian Lapuschkin

Anna Hedström, Philine Bommer, Kristoffer K. Wickstrøm et al.

One of the unsolved challenges in the field of Explainable AI (XAI) is determining how to most reliably estimate the quality of an explanation method in the absence of ground truth explanation labels. Resolving this issue is of utmost importance as the evaluation outcomes generated by competing evaluation methods (or ''quality estimators''), which aim at measuring the same property of an explanation method, frequently present conflicting rankings. Such disagreements can be challenging for practitioners to interpret, thereby complicating their ability to select the best-performing explanation method. We address this problem through a meta-evaluation of different quality estimators in XAI, which we define as ''the process of evaluating the evaluation method''. Our novel framework, MetaQuantus, analyses two complementary performance characteristics of a quality estimator: its resilience to noise and reactivity to randomness, thus circumventing the need for ground truth labels. We demonstrate the effectiveness of our framework through a series of experiments, targeting various open questions in XAI such as the selection and hyperparameter optimisation of quality estimators. Our work is released under an open-source license (https://github.com/annahedstroem/MetaQuantus) to serve as a development tool for XAI- and Machine Learning (ML) practitioners to verify and benchmark newly constructed quality estimators in a given explainability context. With this work, we provide the community with clear and theoretically-grounded guidance for identifying reliable evaluation methods, thus facilitating reproducibility in the field.

Frederik Pahde, Maximilian Dreyer, Wojciech Samek et al.

State-of-the-art machine learning models often learn spurious correlations embedded in the training data. This poses risks when deploying these models for high-stake decision-making, such as in medical applications like skin cancer detection. To tackle this problem, we propose Reveal to Revise (R2R), a framework entailing the entire eXplainable Artificial Intelligence (XAI) life cycle, enabling practitioners to iteratively identify, mitigate, and (re-)evaluate spurious model behavior with a minimal amount of human interaction. In the first step (1), R2R reveals model weaknesses by finding outliers in attributions or through inspection of latent concepts learned by the model. Secondly (2), the responsible artifacts are detected and spatially localized in the input data, which is then leveraged to (3) revise the model behavior. Concretely, we apply the methods of RRR, CDEP and ClArC for model correction, and (4) (re-)evaluate the model's performance and remaining sensitivity towards the artifact. Using two medical benchmark datasets for Melanoma detection and bone age estimation, we apply our R2R framework to VGG, ResNet and EfficientNet architectures and thereby reveal and correct real dataset-intrinsic artifacts, as well as synthetic variants in a controlled setting. Completing the XAI life cycle, we demonstrate multiple R2R iterations to mitigate different biases. Code is available on https://github.com/maxdreyer/Reveal2Revise.

Maximilian Dreyer, Reduan Achtibat, Wojciech Samek et al.

Ensuring both transparency and safety is critical when deploying Deep Neural Networks (DNNs) in high-risk applications, such as medicine. The field of explainable AI (XAI) has proposed various methods to comprehend the decision-making processes of opaque DNNs. However, only few XAI methods are suitable of ensuring safety in practice as they heavily rely on repeated labor-intensive and possibly biased human assessment. In this work, we present a novel post-hoc concept-based XAI framework that conveys besides instance-wise (local) also class-wise (global) decision-making strategies via prototypes. What sets our approach apart is the combination of local and global strategies, enabling a clearer understanding of the (dis-)similarities in model decisions compared to the expected (prototypical) concept use, ultimately reducing the dependence on human long-term assessment. Quantifying the deviation from prototypical behavior not only allows to associate predictions with specific model sub-strategies but also to detect outlier behavior. As such, our approach constitutes an intuitive and explainable tool for model validation. We demonstrate the effectiveness of our approach in identifying out-of-distribution samples, spurious model behavior and data quality issues across three datasets (ImageNet, CUB-200, and CIFAR-10) utilizing VGG, ResNet, and EfficientNet architectures. Code is available on https://github.com/maxdreyer/pcx.

Maximilian Dreyer, Frederik Pahde, Christopher J. Anders et al.

Deep Neural Networks are prone to learning spurious correlations embedded in the training data, leading to potentially biased predictions. This poses risks when deploying these models for high-stake decision-making, such as in medical applications. Current methods for post-hoc model correction either require input-level annotations which are only possible for spatially localized biases, or augment the latent feature space, thereby hoping to enforce the right reasons. We present a novel method for model correction on the concept level that explicitly reduces model sensitivity towards biases via gradient penalization. When modeling biases via Concept Activation Vectors, we highlight the importance of choosing robust directions, as traditional regression-based approaches such as Support Vector Machines tend to result in diverging directions. We effectively mitigate biases in controlled and real-world settings on the ISIC, Bone Age, ImageNet and CelebA datasets using VGG, ResNet and EfficientNet architectures. Code is available on https://github.com/frederikpahde/rrclarc.

Leander Weber, Jim Berend, Moritz Weckbecker et al.

Gradient-based optimization has been a cornerstone of machine learning that enabled the vast advances of Artificial Intelligence (AI) development over the past decades. However, this type of optimization requires differentiation, and with recent evidence of the benefits of non-differentiable (e.g. neuromorphic) architectures over classical models w.r.t. efficiency, such constraints can become limiting in the future. We present Layer-wise Feedback Propagation (LFP), a novel training principle for neural network-like predictors that utilizes methods from the domain of explainability to decompose a reward to individual neurons based on their respective contributions. Leveraging these neuron-wise rewards, our method then implements a greedy approach reinforcing helpful parts of the network and weakening harmful ones. While having comparable computational complexity to gradient descent, LFP does not require gradient computation and generates sparse and thereby memory- and energy-efficient parameter updates and models. We establish the convergence of LFP theoretically and empirically, demonstrating its effectiveness on various models and datasets. Via two applications - neural network pruning and the approximation-free training of Spiking Neural Networks (SNNs) - we demonstrate that LFP combines increased efficiency in terms of computation and representation with flexibility w.r.t. choice of model architecture and objective function. Our code is available at https://github.com/leanderweber/layerwise-feedback-propagation.

Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Reduan Achtibat et al.

To solve ever more complex problems, Deep Neural Networks are scaled to billions of parameters, leading to huge computational costs. An effective approach to reduce computational requirements and increase efficiency is to prune unnecessary components of these often over-parameterized networks. Previous work has shown that attribution methods from the field of eXplainable AI serve as effective means to extract and prune the least relevant network components in a few-shot fashion. We extend the current state by proposing to explicitly optimize hyperparameters of attribution methods for the task of pruning, and further include transformer-based networks in our analysis. Our approach yields higher model compression rates of large transformer- and convolutional architectures (VGG, ResNet, ViT) compared to previous works, while still attaining high performance on ImageNet classification tasks. Here, our experiments indicate that transformers have a higher degree of over-parameterization compared to convolutional neural networks. Code is available at https://github.com/erfanhatefi/Pruning-by-eXplaining-in-PyTorch.

Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun et al.

The field of eXplainable Artificial Intelligence (XAI) aims to bring transparency to today's powerful but opaque deep learning models. While local XAI methods explain individual predictions in form of attribution maps, thereby identifying where important features occur (but not providing information about what they represent), global explanation techniques visualize what concepts a model has generally learned to encode. Both types of methods thus only provide partial insights and leave the burden of interpreting the model's reasoning to the user. In this work we introduce the Concept Relevance Propagation (CRP) approach, which combines the local and global perspectives and thus allows answering both the "where" and "what" questions for individual predictions. We demonstrate the capability of our method in various settings, showcasing that CRP leads to more human interpretable explanations and provides deep insights into the model's representation and reasoning through concept atlases, concept composition analyses, and quantitative investigations of concept subspaces and their role in fine-grained decision making.

Minh An Pham, Anton Segeler, Thomas Wiegand et al.

Function vectors (FVs) are task representations elicited during in-context learning that can be used to steer Large Language Models (LLMs). However, design choices in their formulation remain underexplored. In this work, we study the impact of varying FV definitions for instructions along two degrees of freedom: attention head selection and steering. For head selection, using gradient-based attributions with Layer-wise Relevance Propagation (LRP) substantially improves efficiency as well as accuracy. For FV steering, applying it in a distributed manner yields a higher accuracy compared to simple aggregation. Our code is publicly available.

Leander Weber, Sebastian Lapuschkin, Alexander Binder et al.

Explainable Artificial Intelligence (XAI) is an emerging research field bringing transparency to highly complex and opaque machine learning (ML) models. Despite the development of a multitude of methods to explain the decisions of black-box classifiers in recent years, these tools are seldomly used beyond visualization purposes. Only recently, researchers have started to employ explanations in practice to actually improve models. This paper offers a comprehensive overview over techniques that apply XAI practically for improving various properties of ML models, and systematically categorizes these approaches, comparing their respective strengths and weaknesses. We provide a theoretical perspective on these methods, and show empirically through experiments on toy and realistic settings how explanations can help improve properties such as model generalization ability or reasoning, among others. We further discuss potential caveats and drawbacks of these methods. We conclude that while model improvement based on XAI can have significant beneficial effects even on complex and not easily quantifyable model properties, these methods need to be applied carefully, since their success can vary depending on a multitude of factors, such as the model and dataset used, or the employed explanation method.

Johanna Vielhaben, Sebastian Lapuschkin, Grégoire Montavon et al.

The field of eXplainable Artificial Intelligence (XAI) has greatly advanced in recent years, but progress has mainly been made in computer vision and natural language processing. For time series, where the input is often not interpretable, only limited research on XAI is available. In this work, we put forward a virtual inspection layer, that transforms the time series to an interpretable representation and allows to propagate relevance attributions to this representation via local XAI methods like layer-wise relevance propagation (LRP). In this way, we extend the applicability of a family of XAI methods to domains (e.g. speech) where the input is only interpretable after a transformation. Here, we focus on the Fourier transformation which is prominently applied in the interpretation of time series and LRP and refer to our method as DFT-LRP. We demonstrate the usefulness of DFT-LRP in various time series classification settings like audio and electronic health records. We showcase how DFT-LRP reveals differences in the classification strategies of models trained in different domains (e.g., time vs. frequency domain) or helps to discover how models act on spurious correlations in the data.

Alexander Binder, Leander Weber, Sebastian Lapuschkin et al.

While the evaluation of explanations is an important step towards trustworthy models, it needs to be done carefully, and the employed metrics need to be well-understood. Specifically model randomization testing is often overestimated and regarded as a sole criterion for selecting or discarding certain explanation methods. To address shortcomings of this test, we start by observing an experimental gap in the ranking of explanation methods between randomization-based sanity checks [1] and model output faithfulness measures (e.g. [25]). We identify limitations of model-randomization-based sanity checks for the purpose of evaluating explanations. Firstly, we show that uninformative attribution maps created with zero pixel-wise covariance easily achieve high scores in this type of checks. Secondly, we show that top-down model randomization preserves scales of forward pass activations with high probability. That is, channels with large activations have a high probility to contribute strongly to the output, even after randomization of the network on top of them. Hence, explanations after randomization can only be expected to differ to a certain extent. This explains the observed experimental gap. In summary, these results demonstrate the inadequacy of model-randomization-based sanity checks as a criterion to rank attribution methods.

Maximilian Dreyer, Reduan Achtibat, Thomas Wiegand et al.

Applying traditional post-hoc attribution methods to segmentation or object detection predictors offers only limited insights, as the obtained feature attribution maps at input level typically resemble the models' predicted segmentation mask or bounding box. In this work, we address the need for more informative explanations for these predictors by proposing the post-hoc eXplainable Artificial Intelligence method L-CRP to generate explanations that automatically identify and visualize relevant concepts learned, recognized and used by the model during inference as well as precisely locate them in input space. Our method therefore goes beyond singular input-level attribution maps and, as an approach based on the recently published Concept Relevance Propagation technique, is efficiently applicable to state-of-the-art black-box architectures in segmentation and object detection, such as DeepLabV3+ and YOLOv6, among others. We verify the faithfulness of our proposed technique by quantitatively comparing different concept attribution methods, and discuss the effect on explanation complexity on popular datasets such as CityScapes, Pascal VOC and MS COCO 2017. The ability to precisely locate and communicate concepts is used to reveal and verify the use of background features, thereby highlighting possible biases of the model.

Samar Heydari, Jawher Said, Galip Ümit Yolcu et al. · mit

Deep learning models for flood and wildfire segmentation and object detection enable precise, real-time disaster localization when deployed on embedded drone platforms. However, in natural disaster management, the lack of transparency in their decision-making process hinders human trust required for emergency response. To address this, we present an explainability framework for understanding flood segmentation and car detection predictions on the widely used PIDNet and YOLO architectures. More specifically, we introduce a novel redistribution strategy that extends Layer-wise Relevance Propagation (LRP) explanations for sigmoid-gated element-wise fusion layers. This extension allows LRP relevances to flow through the fusion modules of PIDNet, covering the entire computation graph back to the input image. Furthermore, we apply Prototypical Concept-based Explanations (PCX) to provide both local and global explanations at the concept level, revealing which learned features drive the segmentation and detection of specific disaster semantic classes. Experiments on a publicly available flood dataset show that our framework provides reliable and interpretable explanations while maintaining near real-time inference capabilities, rendering it suitable for deployment on resource-constrained platforms, such as Unmanned Aerial Vehicles (UAVs).

Gabriel Nobis, Maximilian Springenberg, Marco Aversa et al.

We introduce the first continuous-time score-based generative model that leverages fractional diffusion processes for its underlying dynamics. Although diffusion models have excelled at capturing data distributions, they still suffer from various limitations such as slow convergence, mode-collapse on imbalanced data, and lack of diversity. These issues are partially linked to the use of light-tailed Brownian motion (BM) with independent increments. In this paper, we replace BM with an approximation of its non-Markovian counterpart, fractional Brownian motion (fBM), characterized by correlated increments and Hurst index $H \in (0,1)$, where $H=0.5$ recovers the classical BM. To ensure tractable inference and learning, we employ a recently popularized Markov approximation of fBM (MA-fBM) and derive its reverse-time model, resulting in generative fractional diffusion models (GFDM). We characterize the forward dynamics using a continuous reparameterization trick and propose augmented score matching to efficiently learn the score function, which is partly known in closed form, at minimal added cost. The ability to drive our diffusion model via MA-fBM offers flexibility and control. $H \leq 0.5$ enters the regime of rough paths whereas $H>0.5$ regularizes diffusion paths and invokes long-term memory. The Markov approximation allows added control by varying the number of Markov processes linearly combined to approximate fBM. Our evaluations on real image datasets demonstrate that GFDM achieves greater pixel-wise diversity and enhanced image quality, as indicated by a lower FID, offering a promising alternative to traditional diffusion models

Jonas R. Naujoks, Aleksander Krasowski, Moritz Weckbecker et al.

Recently, physics-informed neural networks (PINNs) have emerged as a flexible and promising application of deep learning to partial differential equations in the physical sciences. While offering strong performance and competitive inference speeds on forward and inverse problems, their black-box nature limits interpretability, particularly regarding alignment with expected physical behavior. In the present work, we explore the application of influence functions (IFs) to validate and debug PINNs post-hoc. Specifically, we apply variations of IF-based indicators to gauge the influence of different types of collocation points on the prediction of PINNs applied to a 2D Navier-Stokes fluid flow problem. Our results demonstrate how IFs can be adapted to PINNs to reveal the potential for further studies. The code is publicly available at https://github.com/aleks-krasowski/PINNfluence.

Frederik Pahde, Galip Ümit Yolcu, Alexander Binder et al.

Explainable AI (XAI) is slowly becoming a key component for many AI applications. Rule-based and modified backpropagation XAI approaches however often face challenges when being applied to modern model architectures including innovative layer building blocks, which is caused by two reasons. Firstly, the high flexibility of rule-based XAI methods leads to numerous potential parameterizations. Secondly, many XAI methods break the implementation-invariance axiom because they struggle with certain model components, e.g., BatchNorm layers. The latter can be addressed with model canonization, which is the process of re-structuring the model to disregard problematic components without changing the underlying function. While model canonization is straightforward for simple architectures (e.g., VGG, ResNet), it can be challenging for more complex and highly interconnected models (e.g., DenseNet). Moreover, there is only little quantifiable evidence that model canonization is beneficial for XAI. In this work, we propose canonizations for currently relevant model blocks applicable to popular deep neural network architectures,including VGG, ResNet, EfficientNet, DenseNets, as well as Relation Networks. We further suggest a XAI evaluation framework with which we quantify and compare the effect sof model canonization for various XAI methods in image classification tasks on the Pascal-VOC and ILSVRC2017 datasets, as well as for Visual Question Answering using CLEVR-XAI. Moreover, addressing the former issue outlined above, we demonstrate how our evaluation framework can be applied to perform hyperparameter search for XAI methods to optimize the quality of explanations.

Sami Ede, Serop Baghdadlian, Leander Weber et al.

The ability to continuously process and retain new information like we do naturally as humans is a feat that is highly sought after when training neural networks. Unfortunately, the traditional optimization algorithms often require large amounts of data available during training time and updates wrt. new data are difficult after the training process has been completed. In fact, when new data or tasks arise, previous progress may be lost as neural networks are prone to catastrophic forgetting. Catastrophic forgetting describes the phenomenon when a neural network completely forgets previous knowledge when given new information. We propose a novel training algorithm called training by explaining in which we leverage Layer-wise Relevance Propagation in order to retain the information a neural network has already learned in previous tasks when training on new data. The method is evaluated on a range of benchmark datasets as well as more complex data. Our method not only successfully retains the knowledge of old tasks within the neural networks but does so more resource-efficiently than other state-of-the-art solutions.

Fabian Horst, Djordje Slijepcevic, Matthias Zeppelzauer et al.

State-of-the-art machine learning (ML) models are highly effective in classifying gait analysis data, however, they lack in providing explanations for their predictions. This "black-box" characteristic makes it impossible to understand on which input patterns, ML models base their predictions. The present study investigates whether Explainable Artificial Intelligence methods, i.e., Layer-wise Relevance Propagation (LRP), can be useful to enhance the explainability of ML predictions in gait classification. The research question was: Which input patterns are most relevant for an automated gender classification model and do they correspond to characteristics identified in the literature? We utilized a subset of the GAITREC dataset containing five bilateral ground reaction force (GRF) recordings per person during barefoot walking of 62 healthy participants: 34 females and 28 males. Each input signal (right and left side) was min-max normalized before concatenation and fed into a multi-layer Convolutional Neural Network (CNN). The classification accuracy was obtained over a stratified ten-fold cross-validation. To identify gender-specific patterns, the input relevance scores were derived using LRP. The mean classification accuracy of the CNN with 83.3% showed a clear superiority over the zero-rule baseline of 54.8%.

Djordje Slijepcevic, Fabian Horst, Marvin Simak et al.

Machine learning (ML) models have proven effective in classifying gait analysis data, e.g., binary classification of young vs. older adults. ML models, however, lack in providing human understandable explanations for their predictions. This "black-box" behavior impedes the understanding of which input features the model predictions are based on. We investigated an Explainable Artificial Intelligence method, i.e., Layer-wise Relevance Propagation (LRP), for gait analysis data. The research question was: Which input features are used by ML models to classify age-related differences in walking patterns? We utilized a subset of the AIST Gait Database 2019 containing five bilateral ground reaction force (GRF) recordings per person during barefoot walking of healthy participants. Each input signal was min-max normalized before concatenation and fed into a Convolutional Neural Network (CNN). Participants were divided into three age groups: young (20-39 years), middle-aged (40-64 years), and older (65-79 years) adults. The classification accuracy and relevance scores (derived using LRP) were averaged over a stratified ten-fold cross-validation. The mean classification accuracy of 60.1% was clearly higher than the zero-rule baseline of 37.3%. The confusion matrix shows that the CNN distinguished younger and older adults well, but had difficulty modeling the middle-aged adults.

Annika Frommholz, Fabian Seipel, Sebastian Lapuschkin et al.

Deep neural networks are a promising tool for Audio Event Classification. In contrast to other data like natural images, there are many sensible and non-obvious representations for audio data, which could serve as input to these models. Due to their black-box nature, the effect of different input representations has so far mostly been investigated by measuring classification performance. In this work, we leverage eXplainable AI (XAI), to understand the underlying classification strategies of models trained on different input representations. Specifically, we compare two model architectures with regard to relevant input features used for Audio Event Detection: one directly processes the signal as the raw waveform, and the other takes in its time-frequency spectrogram representation. We show how relevance heatmaps obtained via "Siren"{Layer-wise Relevance Propagation} uncover representation-dependent decision strategies. With these insights, we can make a well-informed decision about the best input representation in terms of robustness and representativity and confirm that the model's classification strategies align with human requirements.

Daniel G. Krakowczyk, Paul Prasse, David R. Reich et al.

Recent work in XAI for eye tracking data has evaluated the suitability of feature attribution methods to explain the output of deep neural sequence models for the task of oculomotric biometric identification. These methods provide saliency maps to highlight important input features of a specific eye gaze sequence. However, to date, its localization analysis has been lacking a quantitative approach across entire datasets. In this work, we employ established gaze event detection algorithms for fixations and saccades and quantitatively evaluate the impact of these events by determining their concept influence. Input features that belong to saccades are shown to be substantially more important than features that belong to fixations. By dissecting saccade events into sub-events, we are able to show that gaze samples that are close to the saccadic peak velocity are most influential. We further investigate the effect of event properties like saccadic amplitude or fixational dispersion on the resulting concept influence.

Oussama Bouanani, Jim Berend, Wojciech Samek et al.

Neuron labeling assigns textual descriptions to internal units of deep networks. Existing approaches typically rely on highly activating examples, often yielding broad or misleading labels by focusing on dominant but incidental visual factors. Prior work such as FALCON introduced contrastive examples -- inputs that are semantically similar to activating examples but elicit low activations -- to sharpen explanations, but it primarily addresses subspace-level interpretability rather than scalable neuron-level labeling. We revisit contrastive explanations for neuron-level labeling in two stages: (1) candidate label generation with vision language models (VLMs) and (2) label assignment with CLIP-like encoders. First, we show that providing contrastive image sets to VLMs yields candidate labels that are more specific and more faithful. Second, we introduce Contrastive Semantic Projection (CSP), an extension of SemanticLens that incorporates contrastive examples directly into its CLIP-based scoring and selection pipeline. Across extensive experiments and a case study on melanoma detection, contrastive labeling improves both faithfulness and semantic granularity over state-of-the-art baselines. Our results demonstrate that contrastive examples are a simple yet powerful and currently underutilized component of neuron labeling and analysis pipelines.

Karam Dawoud, Wojciech Samek, Peter Eisert et al.

In the ever-evolving field of Artificial Intelligence, a critical challenge has been to decipher the decision-making processes within the so-called "black boxes" in deep learning. Over recent years, a plethora of methods have emerged, dedicated to explaining decisions across diverse tasks. Particularly in tasks like image classification, these methods typically identify and emphasize the pivotal pixels that most influence a classifier's prediction. Interestingly, this approach mirrors human behavior: when asked to explain our rationale for classifying an image, we often point to the most salient features or aspects. Capitalizing on this parallel, our research embarked on a user-centric study. We sought to objectively measure the interpretability of three leading explanation methods: (1) Prototypical Part Network, (2) Occlusion, and (3) Layer-wise Relevance Propagation. Intriguingly, our results highlight that while the regions spotlighted by these methods can vary widely, they all offer humans a nearly equivalent depth of understanding. This enables users to discern and categorize images efficiently, reinforcing the value of these methods in enhancing AI transparency.

Simon Ostermann, Daniil Gurgurov, Tanja Baeumel et al.

Post-training adaptation of language models is commonly achieved through parameter updates or input-based methods such as fine-tuning, parameter-efficient adaptation, and prompting. In parallel, a growing body of work modifies internal activations at inference time to influence model behavior, an approach known as steering. Despite increasing use, steering is rarely analyzed within the same conceptual framework as established adaptation methods. In this work, we argue that steering should be regarded as a form of model adaptation. We introduce a set of functional criteria for adaptation methods and use them to compare steering approaches with classical alternatives. This analysis positions steering as a distinct adaptation paradigm based on targeted interventions in activation space, enabling local and reversible behavioral change without parameter updates. The resulting framing clarifies how steering relates to existing methods, motivating a unified taxonomy for model adaptation.

Tobias Labarta, Maximilian Dreyer, Katharina Weitz et al.

Explainable AI (XAI) methods reveal which features influence model predictions, yet provide limited means for practitioners to act on these explanations. Activation steering of components identified via XAI offers a path toward actionable explanations, although its practical utility remains understudied. We introduce an interactive workflow combining SAE-based attribution with activation steering for instance-level analysis of concept usage in vision models, implemented as a web-based tool. Based on this workflow, we conduct semi-structured expert interviews (N=8) with debugging tasks on CLIP to investigate how practitioners reason about, trust, and apply activation steering. We find that steering enables a shift from inspection to intervention-based hypothesis testing (8/8 participants), with most grounding trust in observed model responses rather than explanation plausibility alone (6/8). Participants adopted systematic debugging strategies dominated by component suppression (7/8) and highlighted risks including ripple effects and limited generalization of instance-level corrections. Overall, activation steering renders interpretability more actionable while raising important considerations for safe and effective use.

Mohammad Mahdi Mesgari, Jackie Ma, Wojciech Samek et al.

In the evaluation of attribution quality, the quantitative assessment of explanation legibility is particularly difficult, as it is influenced by varying shapes and internal organization of attributions not captured by simple statistics. To address this issue, we introduce Minimum Spanning Tree Compactness (MST-C), a graph-based structural metric that captures higher-order geometric properties of attributions, such as spread and cohesion. These components are combined into a single score that evaluates compactness, favoring attributions with salient points spread across a small area and spatially organized into few but cohesive clusters. We show that MST-C reliably distinguishes between explanation methods, exposes fundamental structural differences between models, and provides a robust, self-contained diagnostic for explanation compactness that complements existing notions of attribution complexity.

Aleksander Krasowski, René P. Klausen, Aycan Celik et al.

Physics-informed neural networks (PINNs) constitute a flexible deep learning approach for solving partial differential equations (PDEs), which model phenomena ranging from heat conduction to quantum mechanical systems. Despite their flexibility, PINNs offer limited insight into how their predictions deviate from the true solution, hindering trust in their prediction quality. We propose a lightweight post-hoc method that addresses this gap by producing pointwise error estimates for PINN predictions, which offer a natural form of explanation for such models, identifying not just whether a prediction is wrong, but where and by how much. For linear partial differential equations, the error between a PINN approximation and the true solution satisfies the same differential operator as the original problem, but driven by the PINN's PDE residual as its source term. We solve this error equation numerically using finite difference methods requiring no knowledge of the true solution. Evaluated on several benchmark PDEs, our method yields accurate error maps at low computational cost, enabling targeted and interpretable validation of PINNs.

Tobias Labarta, Nhi Hoang, Maximilian Dreyer et al.

The explainable AI (XAI) research community has proposed numerous technical methods, yet deploying explainability as systems remains challenging: Interactive explanation systems require both suitable algorithms and system capabilities that maintain explanation usability across repeated queries, evolving models and data, and governance constraints. We argue that operationalizing XAI requires treating explainability as an information systems problem where user interaction demands induce specific system requirements. We introduce X-SYS, a reference architecture for interactive explanation systems, that guides (X)AI researchers, developers and practitioners in connecting interactive explanation user interfaces (XUI) with system capabilities. X-SYS organizes around four quality attributes named STAR (scalability, traceability, responsiveness, and adaptability), and specifies a five-component decomposition (XUI Services, Explanation Services, Model Services, Data Services, Orchestration and Governance). It maps interaction patterns to system capabilities to decouple user interface evolution from backend computation. We implement X-SYS through SemanticLens, a system for semantic search and activation steering in vision-language models. SemanticLens demonstrates how contract-based service boundaries enable independent evolution, offline/online separation ensures responsiveness, and persistent state management supports traceability. Together, this work provides a reusable blueprint and concrete instantiation for interactive explanation systems supporting end-to-end design under operational constraints.

Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer et al.

Large Language Models are prone to biased predictions and hallucinations, underlining the paramount importance of understanding their model-internal reasoning process. However, achieving faithful attributions for the entirety of a black-box transformer model and maintaining computational efficiency is an unsolved challenge. By extending the Layer-wise Relevance Propagation attribution method to handle attention layers, we address these challenges effectively. While partial solutions exist, our method is the first to faithfully and holistically attribute not only input but also latent representations of transformer models with the computational efficiency similar to a single backward pass. Through extensive evaluations against existing methods on LLaMa 2, Mixtral 8x7b, Flan-T5 and vision transformer architectures, we demonstrate that our proposed approach surpasses alternative methods in terms of faithfulness and enables the understanding of latent representations, opening up the door for concept-based explanations. We provide an LRP library at https://github.com/rachtibat/LRP-eXplains-Transformers.

Bruno Puri, Jim Berend, Sebastian Lapuschkin et al.

Interpretability is crucial for building safe, reliable, and controllable language models, yet existing interpretability pipelines remain costly and difficult to scale. Interpreting a new model typically requires costly training of model-specific sparse autoencoders, manual or semi-automated labeling of SAE components, and their subsequent validation. We introduce Atlas-Alignment, a framework for transferring interpretability across language models by aligning unknown latent spaces to a Concept Atlas - a labeled, human-interpretable latent space - using only shared inputs and lightweight representational alignment techniques. Once aligned, this enables two key capabilities in previously opaque models: (1) semantic feature search and retrieval, and (2) steering generation along human-interpretable atlas concepts. Through quantitative and qualitative evaluations, we show that simple representational alignment methods enable robust semantic retrieval and steerable generation without the need for labeled concept data. Atlas-Alignment thus amortizes the cost of explainable AI and mechanistic interpretability: by investing in one high-quality Concept Atlas, we can make many new models transparent and controllable at minimal marginal cost.

Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben et al.

The field of mechanistic interpretability aims to study the role of individual neurons in Deep Neural Networks. Single neurons, however, have the capability to act polysemantically and encode for multiple (unrelated) features, which renders their interpretation difficult. We present a method for disentangling polysemanticity of any Deep Neural Network by decomposing a polysemantic neuron into multiple monosemantic "virtual" neurons. This is achieved by identifying the relevant sub-graph ("circuit") for each "pure" feature. We demonstrate how our approach allows us to find and disentangle various polysemantic units of ResNet models trained on ImageNet. While evaluating feature visualizations using CLIP, our method effectively disentangles representations, improving upon methods based on neuron activations. Our code is available at https://github.com/maxdreyer/PURE.

Maximilian Dreyer, Jim Berend, Tobias Labarta et al.

Unlike human-engineered systems such as aeroplanes, where each component's role and dependencies are well understood, the inner workings of AI models remain largely opaque, hindering verifiability and undermining trust. This paper introduces SemanticLens, a universal explanation method for neural networks that maps hidden knowledge encoded by components (e.g., individual neurons) into the semantically structured, multimodal space of a foundation model such as CLIP. In this space, unique operations become possible, including (i) textual search to identify neurons encoding specific concepts, (ii) systematic analysis and comparison of model representations, (iii) automated labelling of neurons and explanation of their functional roles, and (iv) audits to validate decision-making against requirements. Fully scalable and operating without human input, SemanticLens is shown to be effective for debugging and validation, summarizing model knowledge, aligning reasoning with expectations (e.g., adherence to the ABCDE-rule in melanoma classification), and detecting components tied to spurious correlations and their associated training data. By enabling component-level understanding and validation, the proposed approach helps bridge the "trust gap" between AI models and traditional engineered systems. We provide code for SemanticLens on https://github.com/jim-berend/semanticlens and a demo on https://semanticlens.hhi-research-insights.eu.

Egor Zverev, Evgenii Kortukov, Alexander Panfilov et al.

Despite their remarkable performance, large language models lack elementary safety features, making them susceptible to numerous malicious attacks. In particular, previous work has identified the absence of an intrinsic separation between instructions and data as a root cause of the success of prompt injection attacks. In this work, we propose a new architectural element, ASIDE, that allows language models to clearly separate instructions and data at the level of embeddings. ASIDE applies an orthogonal rotation to the embeddings of data tokens, thus creating clearly distinct representations of instructions and data tokens without introducing any additional parameters. As we demonstrate experimentally across a range of models, instruction-tuning LLMs with ASIDE (1) leads to highly increased instruction-data separation without a loss in model utility and (2) makes the models more robust to prompt injection benchmarks, even without dedicated safety training. Additionally, we provide insights into the mechanism underlying our method through an analysis of the model representations. The source code and training scripts are openly accessible at https://github.com/egozverev/aside.

Sonia Joseph, Praneet Suresh, Lorenz Hufe et al.

Robust tooling and publicly available pre-trained models have helped drive recent advances in mechanistic interpretability for language models. However, similar progress in vision mechanistic interpretability has been hindered by the lack of accessible frameworks and pre-trained weights. We present Prisma (Access the codebase here: https://github.com/Prisma-Multimodal/ViT-Prisma), an open-source framework designed to accelerate vision mechanistic interpretability research, providing a unified toolkit for accessing 75+ vision and video transformers; support for sparse autoencoder (SAE), transcoder, and crosscoder training; a suite of 80+ pre-trained SAE weights; activation caching, circuit analysis tools, and visualization tools; and educational resources. Our analysis reveals surprising findings, including that effective vision SAEs can exhibit substantially lower sparsity patterns than language SAEs, and that in some instances, SAE reconstructions can decrease model loss. Prisma enables new research directions for understanding vision model internals while lowering barriers to entry in this emerging field.

Bruno Puri, Aakriti Jain, Elena Golimblevskaia et al.

Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While this may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing FADE: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for automatically evaluating feature-to-description alignment. FADE evaluates alignment across four key metrics - Clarity, Responsiveness, Purity, and Faithfulness - and systematically quantifies the causes of the misalignment between features and their descriptions. We apply FADE to analyze existing open-source feature descriptions and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release FADE as an open-source package at: https://github.com/brunibrun/FADE

Leila Arras, Bruno Puri, Patrick Kahardipraja et al.

Various XAI attribution methods have been recently proposed for the transformer architecture, allowing for insights into the decision-making process of large language models by assigning importance scores to input tokens and intermediate representations. One class of methods that seems very promising in this direction includes decomposition-based approaches, i.e., XAI-methods that redistribute the model's prediction logit through the network, as this value is directly related to the prediction. In the previous literature we note though that two prominent methods of this category, namely ALTI-Logit and LRP, have not yet been analyzed in juxtaposition and hence we propose to close this gap by conducting a careful quantitative evaluation w.r.t. ground truth annotations on a subject-verb agreement task, as well as various qualitative inspections, using BERT, GPT-2 and LLaMA-3 as a testbed. Along the way we compare and extend the ALTI-Logit and LRP methods, including the recently proposed AttnLRP variant, from an algorithmic and implementation perspective. We further incorporate in our benchmark two widely-used gradient-based attribution techniques. Finally, we make our carefullly constructed benchmark dataset for evaluating attributions on language models, as well as our code, publicly available in order to foster evaluation of XAI-methods on a well-defined common ground.

Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Reduan Achtibat et al.

Large Language Models (LLMs) are central to many contemporary AI applications, yet their extensive parameter counts pose significant challenges for deployment in memory- and compute-constrained environments. Recent works in eXplainable AI (XAI), particularly on attribution methods, suggest that interpretability can also enable model compression by identifying and removing components irrelevant to inference. In this paper, we leverage Layer-wise Relevance Propagation (LRP) to perform attribution-guided pruning of LLMs. While LRP has shown promise in structured pruning for vision models, we extend it to unstructured pruning in LLMs and demonstrate that it can substantially reduce model size with minimal performance loss. Our method is especially effective in extracting task-relevant subgraphs -- so-called ``circuits'' -- which can represent core functions (e.g., indirect object identification). Building on this, we introduce a technique for model correction, by selectively removing circuits responsible for spurious behaviors (e.g., toxic outputs). All in all, we gather these techniques as a uniform holistic framework and showcase its effectiveness and limitations through extensive experiments for compression, circuit discovery and model correction on Llama and OPT models, highlighting its potential for improving both model efficiency and safety. Our code is publicly available at https://github.com/erfanhatefi/SparC3.

Frederik Pahde, Thomas Wiegand, Sebastian Lapuschkin et al.

Deep neural networks are increasingly employed in high-stakes medical applications, despite their tendency for shortcut learning in the presence of spurious correlations, which can have potentially fatal consequences in practice. Whereas a multitude of works address either the detection or mitigation of such shortcut behavior in isolation, the Reveal2Revise approach provides a comprehensive bias mitigation framework combining these steps. However, effectively addressing these biases often requires substantial labeling efforts from domain experts. In this work, we review the steps of the Reveal2Revise framework and enhance it with semi-automated interpretability-based bias annotation capabilities. This includes methods for the sample- and feature-level bias annotation, providing valuable information for bias mitigation methods to unlearn the undesired shortcut behavior. We show the applicability of the framework using four medical datasets across two modalities, featuring controlled and real-world spurious correlations caused by data artifacts. We successfully identify and mitigate these biases in VGG16, ResNet50, and contemporary Vision Transformer models, ultimately increasing their robustness and applicability for real-world medical tasks. Our code is available at https://github.com/frederikpahde/medical-ai-safety.

Shreyas Gururaj, Lars Grüne, Wojciech Samek et al.

Overfitting is a well-known issue extending even to state-of-the-art (SOTA) Machine Learning (ML) models, resulting in reduced generalization, and a significant train-test performance gap. Mitigation measures include a combination of dropout, data augmentation, weight decay, and other regularization techniques. Among the various data augmentation strategies, occlusion is a prominent technique that typically focuses on randomly masking regions of the input during training. Most of the existing literature emphasizes randomness in selecting and modifying the input features instead of regions that strongly influence model decisions. We propose Relevance-driven Input Dropout (RelDrop), a novel data augmentation method which selectively occludes the most relevant regions of the input, nudging the model to use other important features in the prediction process, thus improving model generalization through informed regularization. We further conduct qualitative and quantitative analyses to study how Relevance-driven Input Dropout (RelDrop) affects model decision-making. Through a series of experiments on benchmark datasets, we demonstrate that our approach improves robustness towards occlusion, results in models utilizing more features within the region of interest, and boosts inference time generalization performance. Our code is available at https://github.com/Shreyas-Gururaj/LRP_Relevance_Dropout.

Galip Ümit Yolcu, Moritz Weckbecker, Thomas Wiegand et al.

Deep learning models achieve remarkable performance, yet their decision-making processes often remain opaque. In response, the field of eXplainable Artificial Intelligence (XAI) has grown significantly over the last decade, primarily focusing on feature attribution methods. Complementing this perspective, Data Attribution (DA) has emerged as a promising paradigm that shifts the focus from features to data provenance. However, existing DA approaches suffer from prohibitively high computational costs and memory demands. Additionally, current attribution methods exhibit low sparsity, hindering the discovery of decisive patterns in the data. We introduce DualXDA, a framework for sparse, efficient and explainable DA, comprised of two interlinked approaches for Dual Data Attribution (DualDA) and eXplainable Data Attribution (XDA): With DualDA, we propose efficient and effective DA, leveraging Support Vector Machine theory to provide fast and naturally sparse data attributions for AI predictions. We demonstrate that DualDA achieves high attribution quality, excels at solving a series of evaluated downstream tasks, while at the same time improving explanation time by a factor of up to 4,100,000$\times$ compared to the original Influence Functions method, and up to 11,000$\times$ compared to the method's most efficient approximation from literature. We further introduce XDA, a method for enhancing Data Attribution with capabilities from feature attribution methods to explain why training samples are relevant for the prediction of a test sample in terms of impactful features. Taken together, our contributions in DualXDA ultimately point towards a future of eXplainable AI applied at unprecedented scale, enabling transparent, efficient and novel analysis of even the largest neural architectures fostering a new generation of accountable AI systems. Code at https://github.com/gumityolcu/DualXDA.

Anna Hedström, Leander Weber, Dilyara Bareeva et al.

The evaluation of explanation methods is a research topic that has not yet been explored deeply, however, since explainability is supposed to strengthen trust in artificial intelligence, it is necessary to systematically review and compare explanation methods in order to confirm their correctness. Until now, no tool with focus on XAI evaluation exists that exhaustively and speedily allows researchers to evaluate the performance of explanations of neural network predictions. To increase transparency and reproducibility in the field, we therefore built Quantus -- a comprehensive, evaluation toolkit in Python that includes a growing, well-organised collection of evaluation metrics and tutorials for evaluating explainable methods. The toolkit has been thoroughly tested and is available under an open-source license on PyPi (or on https://github.com/understandable-machine-intelligence-lab/Quantus/).

Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek et al.

Cross-domain few-shot classification task (CD-FSC) combines few-shot classification with the requirement to generalize across domains represented by datasets. This setup faces challenges originating from the limited labeled data in each class and, additionally, from the domain shift between training and test sets. In this paper, we introduce a novel training approach for existing FSC models. It leverages on the explanation scores, obtained from existing explanation methods when applied to the predictions of FSC models, computed for intermediate feature maps of the models. Firstly, we tailor the layer-wise relevance propagation (LRP) method to explain the predictions of FSC models. Secondly, we develop a model-agnostic explanation-guided training strategy that dynamically finds and emphasizes the features which are important for the predictions. Our contribution does not target a novel explanation method but lies in a novel application of explanations for the training phase. We show that explanation-guided training effectively improves the model generalization. We observe improved accuracy for three different FSC models: RelationNet, cross attention network, and a graph neural network-based formulation, on five few-shot learning datasets: miniImagenet, CUB, Cars, Places, and Plantae. The source code is available at https://github.com/SunJiamei/few-shot-lrp-guided

Sören Becker, Johanna Vielhaben, Marcel Ackermann et al.

Explainable Artificial Intelligence (XAI) is targeted at understanding how models perform feature selection and derive their classification decisions. This paper explores post-hoc explanations for deep neural networks in the audio domain. Notably, we present a novel Open Source audio dataset consisting of 30,000 audio samples of English spoken digits which we use for classification tasks on spoken digits and speakers' biological sex. We use the popular XAI technique Layer-wise Relevance Propagation (LRP) to identify relevant features for two neural network architectures that process either waveform or spectrogram representations of the data. Based on the relevance scores obtained from LRP, hypotheses about the neural networks' feature selection are derived and subsequently tested through systematic manipulations of the input data. Further, we take a step beyond visual explanations and introduce audible heatmaps. We demonstrate the superior interpretability of audible explanations over visual ones in a human user study.

Michael Gerstenberger, Sebastian Lapuschkin, Peter Eisert et al.

Despite significant advances in machine learning, decision-making of artificial agents is still not perfect and often requires post-hoc human interventions. If the prediction of a model relies on unreasonable factors it is desirable to remove their effect. Deep interactive prototype adjustment enables the user to give hints and correct the model's reasoning. In this paper, we demonstrate that prototypical-part models are well suited for this task as their prediction is based on prototypical image patches that can be interpreted semantically by the user. It shows that even correct classifications can rely on unreasonable prototypes that result from confounding variables in a dataset. Hence, we propose simple yet effective interaction schemes for inference adjustment: The user is consulted interactively to identify faulty prototypes. Non-object prototypes can be removed by prototype masking or a custom mode of deselection training. Interactive prototype rejection allows machine learning naïve users to adjust the logic of reasoning without compromising the accuracy.

Jakob Paul Zimmermann, Jim Berend, Georg Loho et al.

Attribution methods explain which input features drive a model's prediction, making them central to model debugging and mechanistic interpretability. Yet backward attribution methods, including gradients, LRP, and transformer-specific rules, lack a shared framework in which to compare the underlying backward calculations. We introduce such a framework by recasting backward attribution as a two-player game on an extended network graph, building on Gaubert and Vlassopoulos' ReLU Net Game. Gradients and the full alpha-beta-LRP family arise as integrals over game trajectories under specific equilibria, so attribution maps become projections of trajectory distributions rather than the primary object. Desired explanation properties, such as localisation focus, robustness to input noise, or stable attention routing, can be specified as game-theoretic concepts, including policy regularization, risk aversion, and extended action sets, and translate directly into novel adaptations of the well-known backward rules. On ViT-B/16, one such selected adaptation of alpha-beta-LRP outperforms prior transformer-specific backward methods across all considered localisation metrics.

Anna Hedström, Leander Weber, Sebastian Lapuschkin et al.

The Model Parameter Randomisation Test (MPRT) is highly recognised in the eXplainable Artificial Intelligence (XAI) community due to its fundamental evaluative criterion: explanations should be sensitive to the parameters of the model they seek to explain. However, recent studies have raised several methodological concerns for the empirical interpretation of MPRT. In response, we propose two modifications to the original test: Smooth MPRT and Efficient MPRT. The former reduces the impact of noise on evaluation outcomes via sampling, while the latter avoids the need for biased similarity measurements by re-interpreting the test through the increase in explanation complexity after full model randomisation. Our experiments show that these modifications enhance the metric reliability, facilitating a more trustworthy deployment of explanation methods.

Anna Hedström, Leander Weber, Sebastian Lapuschkin et al.

The Model Parameter Randomisation Test (MPRT) is widely acknowledged in the eXplainable Artificial Intelligence (XAI) community for its well-motivated evaluative principle: that the explanation function should be sensitive to changes in the parameters of the model function. However, recent works have identified several methodological caveats for the empirical interpretation of MPRT. To address these caveats, we introduce two adaptations to the original MPRT -- Smooth MPRT and Efficient MPRT, where the former minimises the impact that noise has on the evaluation results through sampling and the latter circumvents the need for biased similarity measurements by re-interpreting the test through the explanation's rise in complexity, after full parameter randomisation. Our experimental results demonstrate that these proposed variants lead to improved metric reliability, thus enabling a more trustworthy application of XAI methods.

Dilyara Bareeva, Maximilian Dreyer, Frederik Pahde et al.

Deep Neural Networks are prone to learning and relying on spurious correlations in the training data, which, for high-risk applications, can have fatal consequences. Various approaches to suppress model reliance on harmful features have been proposed that can be applied post-hoc without additional training. Whereas those methods can be applied with efficiency, they also tend to harm model performance by globally shifting the distribution of latent features. To mitigate unintended overcorrection of model behavior, we propose a reactive approach conditioned on model-derived knowledge and eXplainable Artificial Intelligence (XAI) insights. While the reactive approach can be applied to many post-hoc methods, we demonstrate the incorporation of reactivity in particular for P-ClArC (Projective Class Artifact Compensation), introducing a new method called R-ClArC (Reactive Class Artifact Compensation). Through rigorous experiments in controlled settings (FunnyBirds) and with a real-world dataset (ISIC2019), we show that introducing reactivity can minimize the detrimental effect of the applied correction while simultaneously ensuring low reliance on spurious features.

Dilyara Bareeva, Marina M. -C. Höhne, Alexander Warnecke et al.

Feature Visualization (FV) is a widely used technique for interpreting the concepts learned by Deep Neural Networks (DNNs), which synthesizes input patterns that maximally activate a given feature. Despite its popularity, the trustworthiness of FV explanations has received limited attention. In this paper, we introduce a novel method, Gradient Slingshots, that enables manipulation of FV without modifying the model architecture or significantly degrading its performance. By shaping new trajectories in the off-distribution regions of the activation landscape of a feature, we coerce the optimization process to converge in a predefined visualization. We evaluate our approach on several DNN architectures, demonstrating its ability to replace faithfuls FV with arbitrary targets. These results expose a critical vulnerability: auditors relying solely on FV may accept entirely fabricated explanations. To mitigate this risk, we propose a straightforward defense and quantitatively demonstrate its effectiveness.

Patrick Kahardipraja, Reduan Achtibat, Thomas Wiegand et al.

Large language models are able to exploit in-context learning to access external knowledge beyond their training data through retrieval-augmentation. While promising, its inner workings remain unclear. In this work, we shed light on the mechanism of in-context retrieval augmentation for question answering by viewing a prompt as a composition of informational components. We propose an attribution-based method to identify specialized attention heads, revealing in-context heads that comprehend instructions and retrieve relevant contextual information, and parametric heads that store entities' relational knowledge. To better understand their roles, we extract function vectors and modify their attention weights to show how they can influence the answer generation process. Finally, we leverage the gained insights to trace the sources of knowledge used during inference, paving the way towards more safe and transparent language models.