Continuous electrocardiogram (ECG) monitoring via wearables offers significant potential for early cardiovascular disease (CVD) detection. However, deploying deep learning models for automated analysis in resource-constrained environments faces reliability challenges due to inevitable Out-of-Distribution (OOD) data. OOD inputs, such as unseen pathologies or noisecorrupted signals, often cause erroneous, high-confidence predictions by standard classifiers, compromising patient safety. Existing OOD detection methods either neglect computational constraints or address noise and unseen classes separately. This paper explores Unsupervised Anomaly Detection (UAD) as an independent, upstream filtering mechanism to improve robustness. We benchmark six UAD approaches, including Deep SVDD, reconstruction-based models, Masked Anomaly Detection, normalizing flows, and diffusion models, optimized via Neural Architecture Search (NAS) under strict resource constraints (at most 512k parameters). Evaluation on PTB-XL and BUT QDB datasets assessed detection of OOD CVD classes and signals unsuitable for analysis due to noise. Results show Deep SVDD consistently achieves the best trade-off between detection and efficiency. In a realistic deployment simulation, integrating the optimized Deep SVDD filter with a diagnostic classifier improved accuracy by up to 21 percentage points over a classifier-only baseline. This study demonstrates that optimized UAD filters can safeguard automated ECG analysis, enabling safer, more reliable continuous cardiovascular monitoring on wearables.

Wearable cardiovascular sensor patches promise continuous, unobtrusive monitoring, but their tight energy, memory, and compute budgets make it unclear whether physiological signals should be analyzed on the device or streamed to the cloud for processing. We study this inference-versus-transmission trade-off for a resource-constrained patch that records synchronized electrocardiogram (ECG) and phonocardiogram (PCG) signals. We propose an end-to-end, multi-modal convolutional neural network (CNN) with early fusion that classifies the two modalities directly on the device, without hand-crafted features. Trained and validated on the PhysioNet/Computing in Cardiology Challenge 2016 dataset, the floating-point model attains an accuracy of 0.975, which is competitive with the best reported results. At the same time, it reduces the parameter count and computational cost by approximately three orders of magnitude. We deploy an 8-bit integer version of the model on a microcontroller with an integrated neural processing unit (NPU) and measure its inference energy. We also benchmark the energy required for Bluetooth Low Energy (BLE) communication on a representative evaluation kit across a range of payload sizes. NPU inference consumes approximately one-seventh of the energy required for CPU inference. For realistic per-second payloads, local inference is also several times more energy efficient than continuous raw-data streaming. These results show that on-device intelligence, rather than constant transmission, is the more energy-efficient basis for always-on wearable cardiovascular monitoring at the edge.

De Bruijn tori, or perfect maps, are two-dimensional periodic arrays of letters from a finite alphabet, where each possible pattern of shape (m, n) appears exactly once in a single period. While the existence of certain de Bruijn tori, such as square tori with odd m = n is an element of {3, 5, 7} and even alphabet sizes, remains unresolved, sub-perfect maps are often sufficient in applications like positional coding. These maps capture a large number of patterns, with each appearing at most once. While previous methods for generating such sub-perfect maps cover only a fraction of the possible patterns, we present a construction method for generating almost perfect maps for arbitrary pattern shapes and arbitrary non-prime alphabet sizes, including the above mentioned square tori with odd m = n is an element of {3, 5, 7} as long that the alphabet size is non-prime. This is achieved through the introduction of de Bruijn rings, a minimal-height sub-perfect map and a formalization of the concept of families of almost perfect maps. The generated sub- perfect maps are easily decodable which makes them perfectly suitable for positional coding applications. (c) 2025 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http:// creativecommons.org/licenses/by/4.0/).

Large language models (LLMs) have demonstrated remarkable success in the field of natural language processing (NLP). Despite their origins in NLP, these algorithms possess the theoretical capability to process any data type represented in an NLP-like format. In this study, we use stock data to illustrate three methodologies for processing regression data with LLMs, employing tokenization and contextualized embeddings. By leveraging the well-known LLM algorithm Bidirectional Encoder Representations from Transformers (BERT) [1], we apply quantitative stock price prediction methodologies to predict stock prices and stock price movements, showcasing the versatility and potential of LLMs in financial data analysis.

Despite the widespread research on text categorization in various Natural Language Processing (NLP) domains, there exists a noticeable void concerning its application to financial data. This study addresses this gap by employing pre -trained Bidirectional Encoder Representations from Transformers (BERT) models, line -tuned specifically for the financial domain, to categorize newspaper articles focusing on financial topics. This is the first time that the dataset presented in this paper has been used. Further we evaluate the efficacy- of established models in sentiment prediction using these rather long texts. Finally, we delve into the intricacies of company-specific sentiment and relevance prediction within these articles, acknowledging the prevalence of multiple companies being mentioned in one article, thus contributing to a more nuanced understanding of text analysis in the financial sector.

Capsule networks offer a promising solution in computer vision by addressing the limitations of convolutional neural networks (CNNs), such as data dependency and viewpoint challenges. Unlike CNNs, capsules reduce the need for data augmentation by enhancing generalization from limited training data. We explore capsules from the perspective of information theory, viewing them as Monte Carlo sampled continuous random variables. We use marginal differential entropy to measure the information content of capsules, and relative entropy to model the agreement between lower-level and higher-level capsules. The proposed entropy voting method aims to maximize capsule marginal entropies and to minimize their relative entropy. We show that our approach performs better or comparably against state-of-the-art capsule networks while significantly improving inference time. This research highlights the synergy between capsules and information theory, providing insights into their combined potential.

In accordance with the findings presented in [34], this study examines the applicability of Machine Learning (ML) models and training strategies from the Natural Language Processing (NLP) domain in addressing time series problems, emphasizing the structural and operational aspects of these models and strategies. Recognizing the structural congruence within the data, we opt for Stock Price Prediction (SPP) as the designated domain to assess the transferability of NLP models and strategies. Building upon initial positive outcomes derived from quantitative SPP models in our ongoing research endeavors, we provide a rationale for exploring a range of additional methods and conducting subsequent research experiments. The outlined research aims to elucidate the efficacy of leveraging NLP models and techniques for addressing time series problems exemplified as SPP.

Privacy is a well-researched area in the context of Federated Learning. Typically, ensuring privacy means that individual data used for local training cannot be reconstructed by other local learners or a central server. Thus, it is the individual data points that should be private, but not the entire distribution of locally available data. In many cases, this makes sense because each data point comes from a different individual while all data points originate from a common global distribution. In this position paper, we address a more challenging task where the privacy of each local data distribution must be preserved. This is relevant for use cases where there is a one-to-one mapping from local learners to users, such as when each local learner is part of a personalized assistant on a smartphone. We provide a definition of this problem case, describe the challenges that need to be addressed, and formulate a possible approach to solve the problem.

We introduce Quantitative Market Situation Embeddings (QMSEs), a pioneering artificial intelligence (AI)-driven methodology for encoding distinct temporal segments of stock markets into high-dimensional contextual embeddings exclusively leveraging quantitative stock data. Building upon prior research, we construe quantitative stock data analogously to Natural Language Processing (NLP) data, thereby adopting Doc2Vec methodologies to effectuate the embedding of stock data similar to document-level representations. We ascertain the efficacy of QMSEs in representing market dynamics by assessing their ability to discern various significant economic downturns post-2000, including but not limited to, the events of 9/11, the Subprime Crisis of 2008, and the Covid-induced market disruption. Moreover, we elucidate the practical utility of QMSEs through their application in employing distance metrics to gauge the rarity of market scenarios, serving as a regularizer in the training of quantitative stock AI models. Subsequently, we proceed to assess the algorithmic identification of analogous market conditions, aiming to elucidate their potential implications for future stock movements. Additionally, we demonstrate the efficacy of QMSEs in reducing data requirements for quantitative stock AI models by leveraging them as condensed representations of stock data.

This document describes the content and usage of the code artifact files of the original paper "End-to-End Multi-Modal Tiny-CNN for Cardiovascular Monitoring on Sensor Patches". In that work, we show the feasibility of applying deep learning for the classification of synchronized electrocardiogram and phonocardiogram recordings under very tight resource constraints. Our model employs an early fusion of data and uses convolutional layers to solve the problem of binary classification of anomalies. We use the "training-a" dataset of the Physionet Challenge 2016 database for evaluation. Further, we demonstrate the applicability of our model on edge devices, such as sensor patches, by estimating processor performance, power consumption, and silicon area.

Adversarial examples are samples that are close to benign samples with respect to a distance metric, but misclassified by a neural network. While adversarial perturbations of images are usually computed for RGB images, we propose perturbing straight on JPEG coefficients with the ability to individually control the perturbation applied on each color channel and frequency. We find that perturbation as a function of perceptual distance is most efficient for medium frequencies, especially when JPEG compression is used in defense. Overall, we show that attacks on JPEG coefficients are more efficient than state-of-the-art methods that (mainly) apply their perturbation in RGB pixel space. This is partly due to the use of the YCbCr color space, which allows to perturb luma information exclusively, but also due to perturbing the cosine transform coefficients instead of pixels. Moreover, adversarial training using such JPEG attacks with various frequency weighting vectors results in generally strong robustness against RGB and YCbCr attacks as well.

This research addresses the pitfalls and strategies for machine learning with small data sets in the context of sensor-based fatigue detection. It is shown that many existing studies in this area rely on small data sets and that classification results can vary considerably depending on the evaluation method. Our analysis is based on a study with 46 subjects performing multiple sets of squat exercises in a laboratory setting. Data from ratings of perceived exertion, inertial measurement units, and pose estimation were used to train and compare different classifiers. Our findings suggest that commonly used evaluation methods, such as leave-one-subject-out, should be used with caution and may not lead to generalizable classifiers. Furthermore, challenges related to imbalanced data and oversampling are discussed.

Digital circles and digital discs satisfy many bizarre anisotropic properties, understanding of which is essential for solving various problems in image analysis and computer graphics. In this paper we study the underlying properties of absentee pixels that appear while covering a digital disc with concentric digital circles. We present, for the first time, a mathematical characterization of these pixels based on number theory and digital geometry. Interestingly, the absentees occur in multitude, and we show that their count varies quadratically with the radius. The notion of infimum parabola and supremum parabola has been used to derive the count of these absentees. Using this parabolic characterization, we derive a necessary and sufficient condition for a pixel to be a disc absentee, and obtain the geometric properties of the absentees. An algorithm to locate the absentees is presented. We show that the ratio of the absentee pixels to the total number of disc pixels approaches a constant with increasing radius. Test results have been furnished to substantiate our theoretical findings.

Twodimensional contour reconstruction from a set of points is a very common problem not only in computer vision. I.e. in graph theory one may ask for the minimal spanning tree or the shortest Hamiltonian graph. In psychology the question arises under which circumstances people are able to recognize certain contours given only a few points. In the context of discrete geometry, there exist a lot of algorithms for 2D contour reconstruction from sampling points. Here a commonly addressed problem is to define an algorithm for which it can be proved that the reconstuction result resembles the original contour if this has been sampled according to certain density criteria. Most of these algorithms can not properly deal with background noise like humans can do. This paper gives an overview of the most important algorithms for contour reconstruction and shows that a relatively new algorithm, called ‘cleaned refinement reduction’ is the most robust one with regard to significant background noise and even shows a reconstruction ability being similar to the one of a child at the age of 4.

Volume-based boundary reconstruction processes often have to deal with non-manifold shapes. Even though many reconstruction algorithms have been proposed for non-manifold surfaces, they usually do not preserve topological properties. Only recently, methods were presented which-given a finite set of surface sample points-result in a mesh representation of the original boundary preserving all or certain neighborhood relations, even if the sampling is sparse and highly noise corrupted. We show that the required sampling conditions of the algorithm called "refinement reduction" limit the guaranteed correctness of the outcome to a small class of shapes. We define new locally adaptive sampling conditions that depend on our new pruned medial axis and finally prove without any restriction on shapes that under these new conditions, the result of "refinement reduction" corresponds to a refinement of a topologically correct mesh in cases where the previous sampling criteria failed. Based on our results we propose a new criterion for locally adaptive point set decimation. We also discuss why our sampling conditions can only lead to a refinement of a correct reconstruction but not necessarily to a correct reconstruction itself. (c) 2011 Elsevier B.V. All rights reserved.

A simple method for generating random smooth connected mildly irregular binary shapes is introduced. It consists of 1) taking the Minkowski sum of a closed linear spline with random vertices and of a disk (in other words, joining consecutive randomly generated points with straight lines drawn with a "large ballpoint pen"); 2) applying Gaussian blur with a large blur radius; and 3) thresholding permissively. With very permissive thresholds and moderately large numbers of seed points, this produces fairly natural-looking "random blobs." One can also generate "cartoonish shadows" and "boldface alphabets" with less permissive thresholds and smaller numbers of seed points. Rotation invariant families of shapes can be generated by drawing the spline vertices from rotation invariant distributions. Results obtained with the uniform distribution on the disk and the binormal distribution are presented. They are contrasted to those obtained with the uniform distribution on the square. Drawing random points from a binormal distribution gives a collection of shapes that look natural over a wide range of numbers of seed points. The shapes derived with the uniform distributions, however, are more "interesting." Thresholds close to the most restrictive value yielding an empty shape when there is only one seed point work well. This critical threshold is easy to compute using the drawing software; thresholding more permissively guarantees a nonempty shape. The most restrictive threshold guaranteeing a connected final shape is analytically estimated using the diameter of the "pen nib" and the Gaussian blur sigma. The various bounds are in agreement.

Since segmentation of surface data is of growing interest, many image segmentation algorithms have already been applied to surfaces. Mesh segmentation can not only be based on color and texture but also on shape properties like curvature. Graph Cuts based methods are very popular and also applicable to mesh data having irregular vertex neighborhoods. However, Graph Cuts suffer from problems such as shortcuts, generation of small regions and sensitivity to variations of the boundary indicator function. Moreover, they are often hard to implement and computationally expensive. Here we show that a quite old approach for image segmentation, a fast and extremely simple greedy algorithm, which we call LexicoGraph Cut, avoids the problems mentioned above. In this paper we use this algorithm to segment triangulated surface meshes and we introduce new boundary indicator functions, which combine color and curvature properties. We show that using these boundary indicator functions the LexicoGraph Cut algorithm produces a globally optimal result regarding a lexicographic optimization criterion.

The exact, subpixel watershed algorithm delivers very accurate watershed boundaries based on a spline interpolation, but is slow and only works in 2D. On the other hand, there are very fast pixel watershed algorithms, but they produce errors not only in certain exotic cases, but also in real-world images and even in the most simple scenarios. In this work, we examine closely the source of these errors and propose a new algorithm that is fast, approximates the exact watersheds (with pixel resolution), and can be extended to 3D.

Existing theories on shape digitization impose strong constraints on admissible shapes, and require error-free data. Consequently, these theories are not applicable to most real-world situations. In this paper, we propose a new approach that overcomes many of these limitations. It assumes that segmentation algorithms represent the detected boundary by a set of points whose deviation from the true contours is bounded. Given these error bounds, we reconstruct boundary connectivity by means of Delaunay triangulation and α-shapes. We prove that this procedure is guaranteed to result in topologically correct image segmentations under certain realistic conditions. Experiments on real and synthetic images demonstrate the good performance of the new method and confirm the predictions of our theory.

Ziel dieser Arbeit ist es, ein Verfahren zu entwickeln, anhand dessen handschriftliche chinesische Schriftzeichen aus antiken Manuskripten in ihre einzelnen Striche segmentiert und anschlieyend bezuglich bestimmter Stricheigenschaften wie unter anderem Strichbreite, -lange und -richtung untersucht werden konnen. Diese Informationen konnen als Grundlage fur eine Handschriftanalyse mit dem Ziel der Schreiberidentikation dienen, indem Schreiber anhand von individuellen Merkmalen der Striche ihres Schriftbildes klassiziert werden. Das Vorgehen umfasst dabei mehrere Schritte: zuerst wird die Kontur eines in einemeinzelnen digitalen Bild gespeicherten Schriftzeichens extrahiert. Die so gewonnenen Konturpunkte werden anschlieyend mit Hilfe der Eingeschrankten Delaunay- Triangulierung zu Dreiecken verbunden. Aus der Triangulierung der Form des Schriftzeichens werden im Folgenden Strichkreuzungen und partielle Striche extrahiert. Fur diese Striche werden anschlieyend die oben beschriebenen Eigenschaften wie Lange und Schreibrichtung berechnet. Im abschlieyenden Schritt werden die so extrahierten Striche abhangig von den fur sie berechneten Eigenschaften in mehrere Klassen chinesischer Grundstriche eingeteilt. Aufbauend auf diese Arbeit kann eine Klassizierung von Schreibern entwickelt werden, welche auf den berechneten Stricheigenschaften der extrahierten Grundstriche arbeitet. The aim of this work is to segment handwritten Chinese characters in antique manuscripts into their individual strokes for a subsequent analysis of the extracted strokes. A comparison of the computed stroke features with those of known scribes may then be used for manuscript analysis and scribe identication. The procedure includes the following steps: First, the contour of a Chinese character is detected using a simple contour detection method. Using the Constrained Delaunay Triangulation, individual contour points are then connected to triangles. These triangles are used to identify stroke parts and intersections and to compute features such as length and writing direction. Partial strokes are merged into more complex strokes depending on their individual features. In the nal step, the extracted strokes are classied into several classes of Chinese strokes. Based on this work, scribe identication using the extracted stroke features will be developed.