Qianli Wang, Van Bach Nguyen, Yihong Liu, Fedor Splitt, Nils Feldhus, Christin Seifert, Hinrich Schuetze, Sebastian Möller, Vera Schmitt. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Reliance on spurious features that coincidentally correlate with task labels (i.e., shortcut learning) remains a major barrier to the reliable deployment of machine learning models, particularly in high-stakes domains like medical diagnostics.Moreover, in such settings retraining models or collecting and labeling additional data is often impractical, limiting the applicability of many existing shortcut mitigation methods.In this paper we propose a lightweight framework that leverages sparse autoencoders to disentangle spurious from core features to mitigate shortcut learning.Our approach requires no model retraining and works even when group annotations are scarce or unavailable for certain classes.Results on standard benchmarks demonstrate that, even with as few as 50 labeled examples, reliance on spurious features can be significantly reduced.

Shortcut learning, i.e., a model's reliance on undesired features not directly relevant to the task, is a major challenge that severely limits the applications of machine learning algorithms, particularly when deploying them to assist in making sensitive decisions, such as in medical diagnostics. In this work, we leverage recent advancements in machine learning to create an unsupervised framework that is capable of both detecting and mitigating shortcut learning in transformers. We validate our method on multiple datasets. Results demonstrate that our framework significantly improves both worst-group accuracy (samples misclassified due to shortcuts) and average accuracy, while minimizing human annotation effort. Moreover, we demonstrate that the detected shortcuts are meaningful and informative to human experts, and that our framework is computationally efficient, allowing it to be run on consumer hardware.

Counterfactual examples are widely employed to enhance the performance and robustness of large language models (LLMs) through counterfactual data augmentation (CDA). However, the selection of the judge model used to evaluate label flipping, the primary metric for assessing the validity of generated counterfactuals for CDA, yields inconsistent results. To decipher this, we define four types of relationships between the counterfactual generator and judge models: being the same model, belonging to the same model family, being independent models, and having an distillation relationship. Through extensive experiments involving two state-of-the-art LLM-based methods, three datasets, four generator models, and 15 judge models, complemented by a user study (n = 90), we demonstrate that judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations. Relationships between the generator and judge models, which are closely aligned with the user study for CDA, result in better model performance and robustness. Nevertheless, we find that the gap between the most effective judge models and the results obtained from the user study remains considerably large. This suggests that a fully automated pipeline for CDA may be inadequate and requires human intervention.

Paul Youssef, Zhixue Zhao, Christin Seifert, Jörg Schlötterer. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Paul Youssef, Zhixue Zhao, Jörg Schlötterer, Christin Seifert. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Jan Trienes, Sebastian Joseph, Jörg Schlötterer, Christin Seifert, Kyle Lo, Wei Xu, Byron Wallace, Junyi Jessy Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recent advances in natural language processing (NLP) can be largely attributed to the advent of pre-trained language models such as BERT and RoBERTa. While these models demonstrate remarkable performance on general datasets, they can struggle in specialized domains such as medicine, where unique domain-specific terminologies, domain-specific abbreviations, and varying document structures are common. This paper explores strategies for adapting these models to domain-specific requirements, primarily through continuous pre-training on domain-specific data. We pre-trained several German medical language models on 2.4B tokens derived from translated public English medical data and 3B tokens of German clinical data. The resulting models were evaluated on various German downstream tasks, including named entity recognition (NER), multi-label classification, and extractive question answering. Our results suggest that models augmented by clinical and translation-based pre-training typically outperform general domain models in medical contexts. We conclude that continuous pre-training has demonstrated the ability to match or even exceed the performance of clinical models trained from scratch. Furthermore, pre-training on clinical data or leveraging translated texts have proven to be reliable methods for domain adaptation in medical NLP tasks.

Counterfactual text generation aims to minimally change a text, such that it is classified differently.Assessing progress in method development for counterfactual text generation is hindered by a non-uniform usage of data sets and metrics in related work.We propose CEval, a benchmark for comparing counterfactual text generation methods.CEval unifies counterfactual and text quality metrics, includes common counterfactual datasets with human annotations, standard baselines (MICE, GDBA, CREST) and the open-source language model LLAMA-2.Our experiments found no perfect method for generating counterfactual text.Methods that excel at counterfactual metrics often produce lower-quality text while LLMs with simple prompts generate high-quality text but struggle with counterfactual criteria.By making CEval available as an open-source Python library, we encourage the community to contribute additional methods and maintain consistent evaluation in future work. 1

Interpretable methods based on prototypical patches recognize various components in an image in order to explain their reasoning to humans. However, existing prototype-based methods can learn prototypes that are not in line with human visual perception, i.e., the same prototype can refer to different concepts in the real world, making interpretation not intuitive. Driven by the principle of explainability-by-design, we introduce PIP-Net (Patch-based Intuitive Prototypes Network): an interpretable image classification model that learns prototypical parts in a self-supervised fashion which correlate better with human vision. PIP-Net can be interpreted as a sparse scoring sheet where the presence of a prototypical part in an image adds evidence for a class. The model can also abstain from a decision for out-of-distribution data by saying “I haven't seen this before”. We only use image-level labels and do not rely on any part annotations. PIP-Net is globally interpretable since the set of learned prototypes shows the entire reasoning of the model. A smaller local explanation locates the relevant prototypes in one image. We show that our prototypes correlate with ground-truth object parts, indicating that PIP-Net closes the “semantic gap” between latent space and pixel space. Hence, our PIP-Net with interpretable prototypes enables users to interpret the decision making process in an intuitive, faithful and semantically meaningful way. Code is available at https://github.com/M-Nauta/PIPNet.

The goal of Explainable AI (XAI) is to make the reasoning of a machine learning model accessible to humans, such that users of an AI system can evaluate and judge the underlying model. Due to the blackbox nature of XAI methods it is, however, hard to disentangle the contribution of a model and the explanation method to the final output. It might be unclear on whether an unexpected output is caused by the model or the explanation method. Explanation models, therefore, need to be evaluated in technical (e.g. fidelity to the model) and user-facing (correspondence to domain knowledge) terms. A recent survey has identified 29 different automated approaches to quantitatively evaluate explanations. In this work, we take an additional perspective and analyse which toolkits and data sets are available. We investigate which evaluation metrics are implemented in the toolkits and whether they produce the same results. We find that only a few aspects of explanation quality are currently covered, data sets are rare and evaluation results are not comparable across different toolkits. Our survey can serve as a guide for the XAI community for identifying future directions of research, and most notably, standardisation of evaluation.

Automatically summarizing radiology reports into a concise impression can reduce the manual burden of clinicians and improve the consistency of reporting.Previous work aimed to enhance content selection and factuality through guided abstractive summarization.However, two key issues persist.First, current methods heavily rely on domain-specific resources to extract the guidance signal, limiting their transferability to domains and languages where those resources are unavailable.Second, while automatic metrics like ROUGE show progress, we lack a good understanding of the errors and failure modes in this task.To bridge these gaps, we first propose a domain-agnostic guidance signal in form of variable-length extractive summaries.Our empirical results on two English benchmarks demonstrate that this guidance signal improves upon unguided summarization while being competitive with domain-specific methods.Additionally, we run an expert evaluation of four systems according to a taxonomy of 11 fine-grained errors.We find that the most pressing differences between automatic summaries and those of radiologists relate to content selection including omissions (up to 52%) and additions (up to 57%).We hypothesize that latent reporting factors and corpus-level inconsistencies may limit models to reliably learn content selection from the available data, presenting promising directions for future work.Incorrect location of a finding?Incorrect severity of a finding?Any other error?Please describe...

Supervised machine learning utilizes large datasets, often with ground truth labels annotated by humans. While some data points are easy to classify, others are hard to classify, which reduces the inter-annotator agreement. This causes noise for the classifier and might affect the user’s perception of the classifier’s performance. In our research, we investigated whether the classification difficulty of a data point influences how strongly a prediction mistake reduces the “perceived accuracy”. In an experimental online study, 225 participants interacted with three fictive classifiers with equal accuracy (73%). The classifiers made prediction mistakes on three different types of data points (easy, difficult, impossible). After the interaction, participants judged the classifier’s accuracy. We found that not all prediction mistakes reduced the perceived accuracy equally. Furthermore, the perceived accuracy differed significantly from the calculated accuracy. To conclude, accuracy and related measures seem unsuitable to represent how users perceive the performance of classifiers.

Simulated user retrieval system interactions enable studies with controlled user behavior. To this end, the SimIIR framework offers static, rule-based methods. We present an extended SimIIR 2.0 version with new components for dynamic user type-specific Markov model-based interactions and more realistic query generation. A flexible modularization ensures that the SimIIR 2.0 framework can serve as a platform to implement, combine, and run the growing number of proposed search behavior and query simulation ideas.

Proximal femur fractures represent a major health concern, and substantially contribute to the morbidity of elderly. Correct classification and diagnosis of hip fractures has a significant impact on mortality, costs and hospital stay. In this paper, we present a method and empirical validation for automatic subclassification of proximal femur fractures and Dutch radiological report generation that does not rely on manually curated data. The fracture classification model was trained on 11,000 X-ray images obtained from 5000 electronic health records in a general hospital. To generate the Dutch reports, we first trained an embedding model on 20,000 radiological reports of pelvic region fractures, and used its embeddings in the report generation model. We trained the report generation model on the 5000 radiological reports associated with the fracture cases. Our report generation model is on par with state-of-the-art in terms of BLEU and ROUGE scores. This is promising, because in contrast to those earlier works, our approach does not require manual preprocessing of either images or the reports. This boosts the applicability of automatic clinical report generation in practice. A quantitative and qualitative user study among medical students found no significant difference in provenance of real and generated reports. A qualitative, in-depth clinical relevance study with medical domain experts showed that from a human perspective the quality of the generated reports approximates the quality of the original reports and highlights challenges in creating sufficiently detailed and versatile training data for automatic radiology report generation.

Automated decision-making systems become increasingly powerful due to higher model complexity. While powerful in prediction accuracy, Deep Learning models are black boxes by nature, preventing users from making informed judgments about the correctness and fairness of such an automated system. Explanations have been proposed as a general remedy to the black box problem. However, it remains unclear if effects of explanations on user trust generalise over varying accuracy levels. In an online user study with 959 participants, we examined the practical consequences of adding explanations for user trust: We evaluated trust for three explanation types on three classifiers of varying accuracy. We find that the influence of our explanations on trust differs depending on the classifier’s accuracy. Thus, the interplay between trust and explanations is more complex than previously reported. Our findings also reveal discrepancies between self-reported and behavioural trust, showing that the choice of trust measure impacts the results.

Prototype-based methods use interpretable representations to address the black-box nature of deep learning models, in contrast to post-hoc explanation methods that only approximate such models. We propose the Neural Prototype Tree (ProtoTree), an intrinsically interpretable deep learning method for fine-grained image recognition. ProtoTree combines prototype learning with decision trees, and thus results in a globally interpretable model by design. Additionally, ProtoTree can locally explain a single prediction by outlining a decision path through the tree. Each node in our binary tree contains a trainable prototypical part. The presence or absence of this learned prototype in an image determines the routing through a node. Decision making is therefore similar to human reasoning: Does the bird have a red throat? And an elongated beak? Then it’s a hummingbird! We tune the accuracy-interpretability trade-off using ensemble methods, pruning and binarizing. We apply pruning without sacrificing accuracy, resulting in a small tree with only 8 learned prototypes along a path to classify a bird from 200 species. An ensemble of 5 ProtoTrees achieves competitive accuracy on the CUB-200-2011 and Stanford Cars data sets. Code is available at github.com/M-Nauta/ProtoTree.

Sleep scoring is an important step for the detection of sleep disorders and usually performed by visual analysis. Since manual sleep scoring is time consuming, machine-learning based approaches have been proposed. Though efficient, these algorithms are black-box in nature and difficult to interpret by clinicians. In this paper, we propose a deep learning architecture for multi-modal sleep scoring, investigate the model's decision making process, and compare the model's reasoning with the annotation guidelines in the AASM manual. Our architecture, called STQS, uses convolutional neural networks (CNN) to automatically extract spatio-temporal features from 3 modalities (EEG, EOG and EMG), a bidirectional long short-term memory (Bi-LSTM) to extract sequential information, and residual connections to combine spatio-temporal and sequential features. We evaluated our model on two large datasets, obtaining an accuracy of 85% and 77% and a macro F1 score of 79% and 73% on SHHS and an in-house dataset, respectively. We further quantify the contribution of various architectural components and conclude that adding LSTM layers improves performance over a spatio-temporal CNN, while adding residual connections does not. Our interpretability results show that the output of the model is well aligned with AASM guidelines, and therefore, the model's decisions correspond to domain knowledge. We also compare multi-modal models and single-channel models and suggest that future research should focus on improving multi-modal models.

Radiologists are required to write free paper text reports for breast screenings in order to assign cancer diagnoses in a later step. The current procedure requires considerable time and needs efficiency. In this paper, to streamline the writing process and keep up with the specific vocabulary, a word prediction tool using neural language models was developed. Consequently, challenges as different languages (English, Dutch), small data sizes and low computational power have been overcome by introducing a novel English-Dutch Radiology Language Modelling process. After defining model architectures, the process involves data preparation, bilevel hyperparameters optimization, configuration transfer and evaluation. The model is able to improve the current workflow and successfully meet the computational constraints, based on both an intrinsic and extrinsic evaluation. Given its flexibility, the model opens the door for future research involving other languages and also an extensive set of real-world applications.

Hospitals often set protocols based on well defined standards to maintain the quality of patient reports. To ensure that the clinicians conform to the protocols, quality assurance of these reports is needed. Patient reports are currently written in free-text format, which complicates the task of quality assurance. In this paper, we present a machine learning based natural language processing system for automatic quality assurance of radiology reports on breast cancer. This is achieved in three steps: we i) identify the top-level structure (headings) of the report, ii) classify the report content into the top-level headings, and iii) convert the free-text detailed findings in the report to a semi-structured format (post-structuring). Top level structure and content of report were predicted with an F1 score of 0.97 and 0.94, respectively, using Support Vector Machine (SVM) classifiers. For automatic structuring, our proposed hierarchical Conditional Random Field (CRF) outperformed the baseline CRF with an F1 score of 0.78 versus 0.71. The determined structure of the report is represented in semi-structured XML format of the free-text report, which helps to easily visualize the conformance of the findings to the protocols. This format also allows easy extraction of specific information for other purposes such as search, evaluation, and research.

Machine learning experts prefer to think of their input as a single, homogeneous, and consistent data set. However, when analyzing large volumes of data, the entire data set may not be manageable on a single server, but must be stored on a distributed file system instead. Moreover, with the pressing demand to deliver explainable models, the experts may no longer focus on the machine learning algorithms in isolation, but must take into account the distributed nature of the data stored, as well as the impact of any data pre-processing steps upstream in their data analysis pipeline. In this paper, we make the point that even basic transformations during data preparation can impact the model learned, and that this is exacerbated in a distributed setting. We then sketch our vision of end-to-end explainability of the model learned, taking the pre-processing into account. In particular, we point out the potentials of linking the contributions of research on data provenance with the efforts on explainability in machine learning. In doing so, we highlight pitfalls we may experience in a distributed system on the way to generating more holistic explanations for our machine learning models.

While deep neural networks (DNNs) have been shown to outperform humans on many vision tasks, their intransparent decision making process inhibits wide-spread uptake, especially in high-risk scenarios. The BagNet architecture was designed to learn visual features that are easier to explain than the feature representation of other convolutional neural networks (CNNs). Previous experiments with BagNet were focused on natural images providing rich texture and color information. In this paper, we investigate the performance and interpretability of BagNet on a data set of human sketches, i.e., a data set with limited color and no texture information. We also introduce a heatmap interpretability score (HI score) to quantify model interpretability and present a user study to examine BagNet interpretability from user perspective. Our results show that BagNet is by far the most interpretable CNN architecture in our experiment setup based on the HI score.

Search experts use advanced query language and search tactics to formulate their queries. However, the effectiveness of those advanced techniques depends on the search engine internals. We propose QueryCrumbs for Experts, a compact visualization, which facilitates insights to the search engine internals and therefore allows the searcher to determine effective search strategies. Treating the search engine as a black box, QueryCrumbs can be seamlessly integrated into existing search interfaces, guiding the user's exploration and assessment of results. QueryCrumbs for Experts visualize the recent search history alongside with a simple and also a qualitative comparison of the result sets, from which conclusions about the search engine internals can be drawn. The evaluation shows that, by identifying specific patterns in the visualization, expert users can gain valuable insights into search engine internals, empowering them to adapt their search accordingly.

In order to realize zero-effort retrieval in a web-context, it is crucial to identify the part of the web page the user is focusing on. In this paper, we investigate the identification of focus paragraphs in web pages. Starting from a naive baseline for paragraph and focus paragraph detection, we conducted an eye-tracking study to evaluate the most promising features. We found that single features (mouse position, paragraph position, mouse activity) are less predictive for gaze which confirms findings from other studies. The results indicate that an algorithm for focus paragraph detection needs to incorporate a weighted combination of those features as well as additional features, e.g. semantic context derived from the user's web history.

Models of human information seeking reveal that search, in particular ad-hoc retrieval, is non-linear and iterative. Despite these findings, todays search user interfaces do not support non-linear navigation, like for example backtracking in time. In this work, we propose QueryCrumbs, a compact and easy-to-understand visualization for navigating the search query history supporting iterative query refinement. We apply a multi-layered interface design to support novices and firsttime users as well as intermediate users. The formative evaluation with first-time and intermediate users showed that the interactions can be easily performed, and the visual encodings were well understood without instructions. Results indicate that QueryCrumbs can support users when searching for information in an iterative manner.

Entity disambiguation is the task of mapping ambiguous terms in natural-language text to its entities in a knowledge base. It finds its application in the extraction of structured data in RDF (Resource Description Framework) from textual documents, but equally so in facilitating artificial intelligence applications, such as Semantic Search, Reasoning and Question & Answering. We propose a new collective, graph-based disambiguation algorithm utilizing semantic entity and document embeddings for robust entity disambiguation. Robust thereby refers to the property of achieving better than state-of-the-art results over a wide range of very different data sets. Our approach is also able to abstain if no appropriate entity can be found for a specific surface form. Our evaluation shows, that our approach achieves significantly (>5%) better results than all other publicly available disambiguation algorithms on 7 of 9 datasets without data set specific tuning. Moreover, we discuss the influence of the quality of the knowledge base on the disambiguation accuracy and indicate that our algorithm achieves better results than non-publicly available state-of-the-art algorithms.

With the availability of free online enrichment services injection of additional, external resources in existing Web content becomes more and more widespread. For the specific area of just-in-time retrieval of digital resources based on web page content, there are no specific guidelines of how to design and integrate the additional user interface components. In this paper, we conceptualise related user interface issues, investigating the central questions: (i) how can a user be visually notified that additional results are available, and (ii) with which user interface elements should the results be presented. Concretely, we identified four different notification styles and six different result presentation styles. In a survey-based study with 75 participants we elicited the users' preferences, revealing a clear preference for the representation style (split pane) and a strong preference for three notification styles (notification bubble, icon appearance and change of icon's appearance). The latter preferences are related to the preferred browser. The results can serve as guideline for designing web-based user interfaces for just-in-time retrieval.

Despite advancing search technologies, information overload has not yet been solved. Getting an overview of information or explorative access to information becomes increasingly difficult with the exponentially increasing amount of information. Search result visualizations, especially for faceted browsing, aim at supporting users to find their way through large document collections. We propose Facets cape, a novel visualization for navigation and refinement of search results allowing users to visually construct complex boolean search queries for narrowing down the search space. This visualization combines Voronoi subdivision and a tag cloud representation of the search facets. Further it includes a preview of action (query preview) and interactions to allow users to focus on important aspects of the data for the task at hand. In a comparative user study with 15 users we compared the visualization to a standard faceted browsing interface for different types of search tasks. The study revealed that participants used the unfamiliar interface as efficiently and effectively as the familiar tree-like display. Results indicate that the Facets cape is a promising way of supporting users in exploring the faceted search space.

User knowledge levels in adaptive learning systems can be assessed based on user interactions that are interpreted as Knowledge Indicating Events (KIE). Such an approach makes complex inferences that may be hard to understand for users, and that are not necessarily accurate. We present MyExperiences, an open learner model designed for showing the users the inferences about them, as well as the underlying data. MyExperiences is one of the first open learner models based on tree maps. It constitutes an example of how research into open learner models and information visualization can be combined in an innovative way.

The advent of consumer-generated and social media has led to a continuous expansion and diversification of the media landscape. Media consumers frequently find themselves assuming the role of media analysts in order to satisfy personal information needs. We propose to employ Knowledge Visualization methods in support of complex media analysis tasks. In this paper, we describe an approach which depicts semantic relationships between key political actors using node-link diagrams. Our contribution comprises a force-directed edge bundling algorithm which accounts for semantic properties of edges, a technical evaluation of the algorithm and a report on a real-world application of the approach. The resulting visualization fosters the identification of high-level edge patterns which indicate strong semantic relationships. It has been published by the Austrian Press Agency APA in 2009.

Classification and categorization are common tasks in data mining and knowledge discovery. Visualizations of classification models can create understanding and trust in data mining models. However, existing visualizations are often complex or restricted to specific classifiers and attributes. In this work, we propose an intuitive visualization system to observe and understand classification processes and results. Our system can handle multiple classes, nominal and numeric attributes, and supports all classifiers whose predictions can be interpreted as probabilities. We state that the possibility to observe the training process of a classifier boosts the understanding of classification results also for non-expert users. In combination with an intuitive visualization, we provide a system to generate in-depth understanding of classification processes and results. Our simulations revealed that the system could support the user to better understand a classifier's decision, and to gain insights into classification processes.

Tag clouds are text-based visual representations of a set of tags usually depicting tag importance by font size. Recent trends in social and collaborative software have greatly increased the popularity of this type of visualization. This paper proposes a family of novel algorithms for tag cloud layout and presents evaluation results obtained from an extensive user study and a technical evaluation. The algorithms address issues found in many common approaches, for example large whitespaces, overlapping tags and restriction to specific boundaries. The layouts computed by these algorithms are compact and clear, have small whitespaces and may feature arbitrary convex polygons as boundaries. The results of the user study and the technical evaluation enable designers to devise a combination of algorithm and parameters which produces satisfying tag cloud layouts for many application scenarios.

We present a computer vision system for the detection and identification of urban objects from mobile phone imagery, e.g., for the application of tourist information services. Recognition is based on MAP decision making over weak object hypotheses from local descriptor responses in the mobile imagery. We present an improvement over the standard SIFT key detector [7] by selecting only informative (i-SIFT) keys for descriptor matching. Selection is applied first to reduce the complexity of the object model and second to accelerate detection by selective filtering. We present results on the MPG-20 mobile phone imagery with severe illumination, scale and viewpoint changes in the images, performing with ≈ 98% accuracy in identification, efficient (100%) background rejection, efficient (0%) false alarm rate, and reliable quality of service under extreme illumination conditions, significantly improving standard SIFT based recognition in every sense, providing - important for mobile vision - runtimes which are ≈ 8 (≈24) times faster for the MPG-20 (ZuBuD) database.

This work provides a framework for learning sequential attention in real-world visual object recognition, using an architecture of three processing stages. The first stage rejects irrelevant local descriptors based on an information theoretic saliency measure, providing candidates for foci of interest (FOI). The second stage investigates the information in the FOI using a codebook matcher and providing weak object hypotheses. The third stage integrates local information via shifts of attention, resulting in chains of descriptor-action pairs that characterize object discrimination. A Q-learner adapts then from explorative search and evaluative feedback from entropy decreases on the attention sequences, eventually prioritizing shifts that lead to a geometry of descriptor-action scanpaths that is highly discriminative with respect to object recognition. The methodology is successfully evaluated on indoors (COIL-20 database) and outdoors (TSG-20 database) imagery, demonstrating significant impact by learning, outperforming standard local descriptor based methods both in recognition accuracy and processing time.

Autonomous mobile agents require object recognition for high level interpretation and localization in complex scenes. In urban environments, recognition of buildings might play a dominant role in robotic systems that need object based navigation, that take advantage of visual feedback and multimodal information for self-localization, or that enable association to related information from the identified semantics. We present a new method – the informative local features approach – based on an information theoretic saliency measure that is rapidly derived from a local Parzen window density estimation in feature subspace. From the learning of a decision tree based mapping to informative features, it enables attentive access to discriminative information and thereby significantly speeds up the recognition process. This approach is highly robust with respect to severe degrees of partial occlusion, noise, and tolerant to some changes in scale and illumination. We present performance evaluation on our publicly available reference object database (TSG-20) that demonstrates the efficiency of this approach, case wise even outperforming the SIFT feature approach [1]. Building recognition will be advantageous in various application domains, such as, mobile mapping, unmanned vehicle navigation, and systems for car driver assistance.