Detecting out-of-distribution (OOD) examples is an important task for deploying reliable machine learning models in safety-critial applications. While post-hoc methods based on the Mahalanobis distance applied to pre-logit features are among the most effective for ImageNet-scale OOD detection, their performance varies significantly across models. We connect this inconsistency to strong variations in feature norms, indicating severe violations of the Gaussian assumption underlying the Mahalanobis distance estimation. We show that simple $\ell_2$-normalization of the features mitigates this problem effectively, aligning better with the premise of normally distributed data with shared covariance matrix. Extensive experiments on 44 models across diverse architectures and pretraining schemes show that $\ell_2$-normalization improves the conventional Mahalanobis distance-based approaches significantly and consistently, and outperforms other recently proposed OOD detection methods.

The ability to paraphrase texts across different complexity levels is essential for creating accessible texts that can be tailored toward diverse reader groups. Thus, we introduce German4All, the first large-scale German dataset of aligned readability-controlled, paragraph-level paraphrases. It spans five readability levels and comprises over 25,000 samples. The dataset is automatically synthesized using GPT-4 and rigorously evaluated through both human and LLM-based judgments. Using German4All, we train an open-source, readability-controlled paraphrasing model that achieves state-of-the-art performance in German text simplification, enabling more nuanced and reader-specific adaptations. We open-source both the dataset and the model to encourage further research on multi-level paraphrasing.

In this paper we present the development of a quality tile detection device for the Internet of Things (IoT) demonstration in a prototype for educational purposes. For this an Arduino UNO microcontroller is used, and the prototype is chosen after an initial design process, together with suitable tile parts, that imitate possible real tiles in a production line in the downscaled prototype. The Artificial Intelligence (AI) task of pattern detection is performed by a Matlab script. Finally, the Arduino, that controls the conveyor belt and the switch of the tile quality detection device is linked to the Matlab Script, that controls the picture detection and AI-script evaluation on the PC and its backcoupling to the Arduino, yielding a semi-automatic tile quality checking procedure with a human in the loop.

Sharpness of minima is a promising quantity that can correlate with generalization in deep networks and, when optimized during training, can improve generalization. However, standard sharpness is not invariant under reparametrizations of neural networks, and, to fix this, reparametrization-invariant sharpness definitions have been proposed, most prominently adaptive sharpness (Kwon et al., 2021). But does it really capture generalization in modern practical settings? We comprehensively explore this question in a detailed study of various definitions of adaptive sharpness in settings ranging from training from scratch on ImageNet and CIFAR-10 to fine-tuning CLIP on ImageNet and BERT on MNLI. We focus mostly on transformers for which little is known in terms of sharpness despite their widespread usage. Overall, we observe that sharpness does not correlate well with generalization but rather with some training parameters like the learning rate that can be positively or negatively correlated with generalization depending on the setup. Interestingly, in multiple cases, we observe a consistent negative correlation of sharpness with OOD generalization implying that sharper minima can generalize better. Finally, we illustrate on a simple model that the right sharpness measure is highly data-dependent, and that we do not understand well this aspect for realistic data distributions.

Out-of-distribution (OOD) detection is the problem of identifying inputs which are unrelated to the in-distribution task. The OOD detection performance when the in-distribution (ID) is ImageNet-1K is commonly being tested on a small range of test OOD datasets. We find that most of the currently used test OOD datasets, including datasets from the open set recognition (OSR) literature, have severe issues: In some cases more than 50$%$ of the dataset contains objects belonging to one of the ID classes. These erroneous samples heavily distort the evaluation of OOD detectors. As a solution, we introduce with NINCO a novel test OOD dataset, each sample checked to be ID free, which with its fine-grained range of OOD classes allows for a detailed analysis of an OOD detector’s strengths and failure modes, particularly when paired with a number of synthetic “OOD unit-tests”. We provide detailed evaluations across a large set of architectures and OOD detection methods on NINCO and the unit-tests, revealing new insights about model weaknesses and the effects of pretraining on OOD detection performance. We provide code and data at https://github.com/j-cb/NINCO.

Sharpness-aware minimization (SAM) was proposed to reduce sharpness of minima and has been shown to enhance generalization performance in various settings. In this work we show that perturbing only the affine normalization parameters (typically comprising 0.1% of the total parameters) in the adversarial step of SAM can outperform perturbing all of the parameters. This finding generalizes to different SAM variants and both ResNet (Batch Normalization) and Vision Transformer (Layer Normalization) architectures. We consider alternative sparse perturbation approaches and find that these do not achieve similar performance enhancement at such extreme sparsity levels, showing that this behaviour is unique to the normalization layers. Although our findings reaffirm the effectiveness of SAM in improving generalization performance, they cast doubt on whether this is solely caused by reduced sharpness.

The field of neuromorphic computing is approaching an era of rapid adoption driven by the urgent need of a substitute for the von Neumann computing architecture. NEUROTEC I: “Neuro-inspired Artificial Intelligence Technologies for the Elec-tronics of the Future” project is an initiative sponsored by the German Federal Ministry of Education and Research (BMBF for its initials in German), that aims to effectively advance the foundations for the utilization and exploitation of neuromorphic computing. NEUROTEC I stands at its successful “final stage” driven by the collaboration from more than 8 institutes from the Jiilich Research Center and the RWTH Aachen University, as well as collaboration from several high-tech industry partners. The NEUROTEC I project considers the field interplay among materials, circuits, design and simulation tools. This paper provides an overview of the project's overall structure and discusses the scientific achievements of its individual activities.

Image-based quality control aims at detecting anomalies in products and is a crucial part of the production process. Challenges arise from the complexity and variety of products, defects, and the rarity of defect occurrence. Quality control thus still relies heavily on manual inspection. Supervised, data driven approaches have greatly improved performance, but suffer from a major drawback: They require large amounts of annotated training data, limiting their economic viability.In this work, we overcome this drawback by leveraging the consistency of defect appearance across fabrics to transfer knowledge about anomalies from one fabric to another. We realize this by adapting the image-to-image translation framework, introducing guidance by means of a segmentation map. We evaluate both image quality as well as usability of generated defects. For the image quality, both classical (L1, L2 and Structured Similarity Measurement (SSIM)) and perceptual (Learned Perceptual Image Patch Similarity (LPIPS)) metrics indicate efficacy of the presented approach, i.e. defect synthesis only occurs in the targeted regions, and the background fabric pattern remains largely unchanged. To evaluate usability of generated defects, we train pseudo-supervised models on synthesized defects of fabrics unseen during training. A comparison with semi-supervised, autoencoder based approaches demonstrates the suitability of our approach, yielding average Area Under the Receiver Operating Characteristic Curve (AUROC) values of 0.81 for pseudo-supervised vs. 0.69 for semi-supervised settings.

The recently finished EU project JuxtaLearn aimed at supporting students’ learning of STEM subjects through the creation, exchange and discussion of learner-made videos. The approach is based on an eight-stage activity cycle in the beginning of which teachers identify specific “stumbling blocks” for a given theme (or “tricky topic”). In JuxtaLearn, video comments were analyzed to extract information on the learners’ acquisition and understanding of domain concepts, especially to detect problems and misconceptions. These analyses were based on mapping texts to networks of concepts (“network-text analysis”) as a basis for further processing. In this article we use data collected from recent field trials to shed light on what is actually discussed when students share their own videos in science domains. Would the aspect of video-making dominate over activities related to a deepening of domain understanding? Our findings indicate that there are different ways of balancing both aspects and interventions will be needed to bring forth the desired blend.

The recent development and distribution of interactive pervasive computing systems, especially those incorporating digital public displays, facilitate a variety of new interactive scenarios in shared public spaces where people gather and spend time to enjoy a social experience. Integrating novel interaction mechanisms related to the sharing and listening of music with digital displays is a good example of the potential of using current technologies to enhance physical and social spaces. This paper explores the design and impact of shared situated interactive systems, combining public display installations with a collaborative music player. To explore the design space of these systems, we conducted two studies involving the deployment of a particular system at different venues. The analysis of our results allowed us to identify specific (group) interactions and events, which were fostered by the system under consideration. Based on these outcomes, we propose a set of features that should foster further engagement with such systems and promote new ways of social interaction.

Recent years have seen a growing interest in supporting learning activities/scenarios that go beyond the traditional classroom context as well as the development of pervasive computing scenarios supported by display installations. In order to explore such interactive scenarios that span video-based learning activities across school contexts, we have developed two web-based functional prototypes of public display applications and performed a field evaluation during an initial test-deployment. The system consists of a public display endpoint providing video content enriched with quizzes related to this content, and a mobile endpoint providing interactivity and user participation. During a three weeks test-deployment at two Swedish schools, the display system was evaluated and important requirements for the next iterations were gathered. This work presents the results of the test-deployment and the users' adoption (usage patterns), and discusses the specialties of introducing such a system into educational environments. The deployment and the corresponding study enabled us to validate in real settings the overall technical approach and to test different perspectives of display usage. The conclusions point to the need to further understand how to promote an integrated view of display utilization in schools.