Cross Entropy

A threshold-based algorithm is presented for the extraction of the pectoral muscle edge in mediolateral oblique view mammograms. The minimum cross-entropy thresholding algorithm is applied to local areas around the pectoral muscle to determine a series of thresholds as a function of area size. Using a model image it is shown that art inflection point in this function corresponds to

Credal predictors are epistemic-uncertainty-aware models that produce a convex set of probabilistic predictions. They provide a principled framework for quantifying predictive epistemic uncertainty (EU) and have been shown to improve model robustness across a range of settings. However, most state-of-the-art (SOTA) methods primarily define EU as disagreement induced by random training initializations, which mainly reflects sensitivity to optimization randomness rather than uncertainty from more substantive sources. In response, we formulate EU as disagreement between models trained under different degrees of relaxation of the i.i.d. assumption between the training and test distributions. Building on this idea, we propose CreDRO, which learns an ensemble of plausible models via distributionally robust optimization. As a result, CreDRO captures EU arising not only from training randomness but also from informative disagreement due to potential train-test distribution shifts. Empirically, CreDRO consistently outperforms SOTA credal approaches on downstream tasks, including outof-distribution detection on extensive benchmarks and selective classification in medical settings.

Informational Approach (IA) to morphology processing has shaped significant body of psycholinguistic research in Serbian and some other languages. However, its central claim that word processing variability is fully accounted for by the amount of information carried by inflected word forms was not systematically tested utilizing full range of the word forms. Three experiments that employed complete Serbian noun form system were conducted, with masculine and neuter gender for the first time. Results obtained generally replicate previously reported, but we demonstrate that lexical-distributional variables, letter string length and form frequency, provide an equal prediction of average lexical decision latencies as it does theoretically driven derived IA measure, bits. Data analysis methods which IA theorizing crucially rests upon are not compatible with experiments' repeated design. Including grammatical complexity as obligatory term into the IA measures proved uncalled for. Models based on summary statistics such as IA hardly provide basis for unequivocal conclusions.

One of the most often-used stopping criteria is the cross-entropy stopping criterion (CESC). The CESC can stop turbo decoder iterations early by calculating mutual information improvements while maintaining bit error rate (BER) performance. Most research on iterative turbo decoding stopping criteria has utilised low-modulation methods, such as binary phase-shift keying. However, a high-speed network requires high modulation to transfer data at high speeds. Hence, a high modulation technique needs to be integrated into the CESC to match its speed. Therefore, the present paper investigated and analysed the effects of the CESC and quadrature amplitude modulation (QAM) on iterative turbo decoding. Three thresholds were simulated and tested under four situations: different code rates, different QAM formats, different code generators, and different frame sizes. The results revealed that in most situations, the use of CESC is suitable only when the signal-to-noise ratio (SNR) is high. This is because the CESC significantly reduces the average iteration number (AIN) while maintaining the BER. The CESC can terminate early at a high SNR and save more than 40% AIN compared with the fixed stopping criterion. Meanwhile, at a low SNR, the CESC fails to terminate early, which results in maximum AIN.

Face recognition is one of the most widely publicized feature in the devices today and hence represents an important problem that should be studied with the utmost priority. As per the recent trends, the Convolutional Neural Network (CNN) based approaches are highly successful in many tasks of Computer Vision including face recognition. The loss function is used on the top of CNN to judge the goodness of any network. In this paper, we present a performance comparison of different loss functions such as Cross-Entropy, Angular Softmax, Additive-Margin Softmax, ArcFace and Marginal Loss for face recognition. The experiments are conducted with two CNN architectures namely, ResNet and Mo-bileNet. Two widely used face datasets namely, CASIA-Webface and MS-Celeb-1M are used for the training and benchmark Labeled Faces in the Wild (LFW) face dataset is used for the testing.

and treatment are crucial factors in preventing extensive retina damage and even vision loss. In today's practice, DR is diagnosed based on the examination of retina's fundus images by experienced medical practitioners based on the presence of three lesion types (i.e. Exudates, Hemorrhages

This paper investigates a new heterogeneous method that dynamically sets the migration period of a distributed Genetic Algorithm (dGA). Each island GA of this multipopulation technique self-adapts the period for exchanging information with the other islands regarding the local evolution process. Thus, the different islands can develop different migration settings behaving like a heterogeneous dGA. The proposed algorithm is tested on a large set of instances of the Max-Cut problem, and it can be easily applied to other optimization problems. The results of this heterogeneous dGA are competitive with the best existing algorithms, with the added advantage of avoiding timeconsuming preliminary tests for tuning the algorithm.

Handwritten Character Recognition (HCR) remains a fundamental yet challenging problem in pattern recognition due to variations in writing styles, distortions, noise, and inter-class similarity. This study proposes an intelligent AI-based framework for robust handwritten character recognition using a Convolutional Neural Network (CNN). The model is trained and evaluated on a self-curated dataset comprising 13,640 grayscale images representing 62character classes, including digits (0-9), lowercase letters (a-z), and uppercase letters (A-Z). Images are standardized to a resolution of 28×28 pixels and normalized to enhance learning efficiency. The proposed CNN architecture leverages hierarchical feature extraction through multiple convolutional and pooling layers, followed by dense layers for classification. Experimental results demonstrate a recognition accuracy of approximately 93%, indicating strong generalization capability despite handwriting variability. The proposed framework emphasizes automated feature learning, eliminating the dependency on handcrafted descriptors traditionally used in character recognition systems. The model exhibits strong adaptability across diverse handwriting patterns, demonstrating robustness to intra-class variations. Furthermore, the lightweight CNN architecture ensures computational efficiency, making the system suitable for real-time applications and deployment in resource-constrained environments. The study also highlights the critical role of dataset quality, preprocessing strategies, and normalization techniques in improving recognition performance. Overall, the findings confirm that deep learning-driven approaches offer a reliable, scalable, and efficient solution for handwritten character recognition.

Brain tumor segmentation is a critical and challenging task in medical image analysis, essential for enhancing patient outcomes through timely detection and treatment. Manual segmentation is both laborious and subjective, highlighting the need for efficient automated solutions. Deep learning, especially models like U-Net, has demonstrated significant potential in biomedical image segmentation. The diversity in brain tumor location, shape, and size adds complexity to automatic segmentation, which involves differentiating abnormal from normal tissues (WM, GM, CSF). Multimodal approaches (e.g., PET/CT, PET/MRI) improve segmentation accuracy by integrating data from various modalities. MRI remains the primary tool for detecting brain tumors, but the process of tumor identification and segmentation is complex and heavily reliant on the expertise of radiologists. Recent advancements in deep learning have markedly enhanced the feasibility of accurate, automated brain tumor segmentation, meeting a critical need in medical imaging. Our proposed model, trained on the TGCA dataset, demonstrated superior performance with DSC 0.94, IoU 0.94 and SI 0.94 compared to traditional models like DeepResnet and Znet. This new method, notable for its performance, success, and pre-processing structure, is highly adaptable for various applications.

We present a novel non-parametric unsupervised segmentation algorithm based on Region Competition ; but implemented within a Level Sets framework . The key novelty of the algorithm is that it can solve N ≥ 2 class segmentation problems using just one embedded surface; this is achieved by controlling the merging and splitting behaviour of the level sets according to a Minimum Description Length (MDL) cost function. This is in contrast to N class region-based Level Set segmentation methods to date which operate by evolving multiple coupled embedded surfaces in parallel . Furthermore, it operates in an unsupervised manner; it is necessary neither to specify the value of N nor the class models a-priori. We argue that the Level Sets methodology provides a more convenient framework for the implementation of the Region Competition algorithm, which is conventionally implemented using region membership arrays due to the lack of a intrinsic curve representation. Finally, we generalise the Gaussian region model used in standard Region Competition to the non-parametric case. The region boundary motion and merge equations become simple expressions containing cross-entropy and entropy terms.

本論文は、AI支援研究における方法論的フレームワークである「プロンプト指向型研究オペレーティングシステム(Research OS)」の構造を、ソフトウェアアーキテクチャ可視化手法であるPlantUML C4モデルを用いて体系的に分析する。C4モデルの4階層(Context/Container/Component/Code)により、Research OSの構造を段階的に可視化し、固定プロンプト層と可変プロンプト層の入れ子構造、10の研究支援モジュールの責務分離、および研究プロセスの時系列展開を明示化した。
システム理論的器楽学への適用事例では、論文執筆期間50%短縮、査読応答時間67%短縮、ジャーナル移行成功率200%向上という定量的効果を実証した。分析から導出された5つの設計原則(制約優先設計、階層的分離、責務の単一性、依存関係の一方向性、検証可能性)は、他の研究領域でResearch OSを実装する際の実践的指針となる。
本研究の学術的貢献は、(1)AI支援研究方法論を「実装可能なシステムアーキテクチャ」として理解する新たな視座の提供、(2)工学的可視化手法の人文学的研究への応用可能性の実証、(3)暗黙知だった研究判断プロセスの形式知化による再現可能性の向上、の3点である。この成果は、デジタル人文学における方法論的多様性の拡張と、科学の科学(Science of Science)における新たな分析対象の提供に寄与する。
キーワード
• プロンプト指向型研究オペレーティングシステム
• PlantUML C4モデル
• AI支援研究方法論
• ソフトウェアアーキテクチャ可視化
• 大規模言語モデル(LLM)
• 固定プロンプト・可変プロンプト
• デジタル人文学
• 研究プロセスの形式知化
• 再現可能性
• システム理論的器楽学
________________________________________
Abstract
This paper systematically analyzes the structure of the "Prompt-Oriented Research Operating System (Research OS)," a methodological framework for AI-assisted research, using the PlantUML C4 model, a software architecture visualization technique. Through the four layers of the C4 model (Context/Container/Component/Code), the structure of the Research OS is progressively visualized, clarifying the nested structure of fixed and variable prompt layers, the responsibility separation of ten research support modules, and the temporal development of research processes.
A case study applying this framework to system-theoretic organology demonstrated quantitative improvements: 50% reduction in paper writing time, 67% reduction in peer review response time, and 200% improvement in journal transfer success rate. The five design principles derived from this analysis (constraint-first design, hierarchical separation, single responsibility, unidirectional dependency, and verifiability) provide practical guidelines for implementing Research OS in other research domains.
The academic contributions of this study are threefold: (1) providing a new perspective for understanding AI-assisted research methodology as "implementable system architecture," (2) demonstrating the applicability of engineering visualization methods to humanistic research, and (3) improving reproducibility through the formalization of previously tacit research decision-making processes. These findings contribute to expanding methodological diversity in digital humanities and providing new analytical objects in the science of science.

Generalized zero shot learning (GZSL) is still a technical challenge of deep learning as it has to recognize both source and target classes without data from target classes. To preserve the semantic relation between source and target classes when only trained with data from source classes, we address the quantification of the knowledge transfer and semantic relation from an informationtheoretic viewpoint. To this end, we follow the prototypical model and format the variables of concern as a probability vector. Leveraging on the proposed probability vector representation, the information measurement such as mutual information and entropy, can be effectively evaluated with simple closed forms. We discuss the choice of common embedding space and distance function when using the prototypical model. Then We propose three information-theoretic loss functions for deterministic GZSL model: a mutual information loss to bridge seen data and target classes; an uncertainty-aware entropy constraint loss to prevent overfitting when using seen data to learn the embedding of target classes; a semantic preserving cross entropy loss to preserve the semantic relation when mapping the semantic representations to the common space. Simulation shows that, as a deterministic model, our proposed method obtains state of the art results on GZSL benchmark datasets. We achieve 21% ∼ 64% improvements over the baseline model -deep calibration network (DCN) and for the first time demonstrate a deterministic model can perform as well as generative ones. Moreover, our proposed model is compatible with generative models. Simulation studies show that by incorporating with f-CLSWGAN, we obtain comparable results compared with advanced generative models.

В роботі запропоновано метод класифікації. Він заснований на комбінованому алгоритмі негативної селекції, який був спочатку розроблений для задач бінарної класифікації. Точність розробленого алгоритму була перевірена експериментальним шляхом з використанням наборів даних мікрочіпів. Експерименти підтвердили, що напрямок змін, внесених в розроблений алгоритм підвищує точність у порівнянні з іншими алгоритмів класифікації. Ключові слова: алгоритм вибору, класифікатор, мікрочіп даних, аналіз головних компонентів, імпульсне перетворення, відбір ознак. In the paper, a classification method is proposed. It is based on Combined Swarm Negative Selection Algorithm, which was originally designed for binary classification problems. The accuracy of developed algorithm was tested in an experimental way with the use of microarray data sets. The experiments confirmed that direction of changes introduced in developed algorithm improves its accuracy in comparison to other classification algorithms.

This paper proposes a data-driven approximate Bayesian computation framework for parameter estimation and uncertainty quantification of epidemic models, which incorporates two novelties: (i) the identification of the initial conditions by using plausible dynamic states that are compatible with observational data; (ii) learning of an informative prior distribution for the model parameters via the cross-entropy method. The new methodology's effectiveness is illustrated with the aid of actual data from the COVID-19 epidemic in Rio de Janeiro city in Brazil, employing an ordinary differential equation-based model with a generalized SEIR mechanistic structure that includes timedependent transmission rate, asymptomatics, and hospitalizations. A minimization problem with two cost terms (number of hospitalizations and deaths) is formulated, and twelve parameters are identified. The calibrated model provides a consistent description of the available data, able to extrapolate forecasts over a few

This paper investigates various combinations of preprocessing methods (attribute selection, normalization, resampling, and imputation) and evaluates their impact on the performance of decision tree models for predicting customer churn. The experiments were performed on the benchmark Cell2Cell dataset due to its ability to address diverse aspects of customer behavior, including value-added services, usage patterns, demographic information, customer service interactions, personal data, and billing data. This comprehensive view of client activities makes it ideal for studying customer churn. The aim of this work is to identify the most effective preprocessing method that can be applied to a real-world telecommunications dataset to improve the effectiveness of customer churn prediction methods. The study systematically examines the effects of imputation methods (K-Nearest Neighbors and statistical imputation), normalization techniques (Median and Median Absolute Deviation Normalization, Min-Max Scaling, and Z-Score Standardization), feature selection using Lasso regression, and resampling using SMOTE Tomek. This results in 16 distinct preprocessed datasets, each reflecting a unique combination of preprocessing steps. An analysis of these datasets was conducted, evaluating the performance metrics of the Decision Tree model on each dataset, including accuracy, precision, recall, F1 score, and ROC-AUC. Key findings highlight that Statistical Imputation, Median and Median Absolute Deviation Normalization, and Lasso feature selection achieved the highest performance, with 0.78 in precision, 0.77 in accuracy, recall, and F1 Score, and 0.74 in ROC-AUC.

Modern machine learning, fueled by large datasets and complex models, faces a critical tension. The statistical principles underpinning learning (generalization, efficiency, robustness) often clash with the computational realities of optimization, especially in a resource constrained environment or when data exhibits inherent geometric structure. This work addresses the theme "Statistics Meets Optimization" by employing an optimization framework explicitly designed to leverage statistical data properties, particularly group invariances/equivariances common in real world data (e.g., spatial rotations in satellite imagery, temporal shifts in sensor data), to achieve significant gains in sample efficiency and convergence speed. We theoretically derive generalization bounds linking the exploitation of data geometry to reduced sample complexity. Empirically, we demonstrate the efficacy of our method on a challenging real world case study, i.e., on predicting crop yield anomalies in Delta State, Nigeria, using limited, noisy, and spatially heterogeneous satellite and meteorological data. Our optimizer achieved a significant performance with 40% less data compared to adaptive baselines (Adam, RMSProp), highlighting the practical impact of statistically-informed optimization, especially for regions facing data scarcity. This work provides a concrete bridge between statistical theory (data structure, efficiency) and optimization practice (algorithm design, scalability), demonstrating that geometry-aware algorithms can democratize effective ML for resource-limited applications.

National accounts statistics are the result of the integration of several data sources. At present, the Italian national accounts use household surveys data for estimating labour units only, not for estimating the monetary aggregates. In this paper we investigate the possibility of using sample data on household income within an independent estimate of the Gross Domestic Product based on the income approach. The aim of this paper is twofold. In the first part we compare household sample data with national accounts trying to assess whether (and to what extent) sample survey data may contribute to the estimate of the Italian GDP. Furthermore we try to actually bring sample income micro data in the estimation process of the Italian NA and measure how the estimate of GDP and its components is affected. Applications are based on the European statistics on income and living conditions and on the Bank of Italy survey on income and wealth.

This report is an output from a project funded by the UK Department for International Development (DFID) under the UK provision of technical assistance to developing countries. The views expressed are not necessarily those of the Department.

Segmenting biomarkers in medical images is crucial for various biotech applications. Despite advances, Transformer and CNN based methods often struggle with variations in staining and morphology, limiting feature extraction. In medical image segmentation, where datasets often have limited sample availability, recent state-of-theart (SOTA) methods achieve higher accuracy by leveraging pre-trained encoders, whereas end-to-end methods tend to underperform. This is due to challenges in effectively transferring rich multiscale features from encoders to decoders, as well as limitations in decoder efficiency. To address these issues, we propose an architecture that captures multi-scale local and global contextual information and a novel decoder design, which effectively integrates features from the encoder, emphasizes important channels and regions, and reconstructs spatial dimensions to enhance segmentation accuracy. Our method, compatible with various encoders, outperforms SOTA methods, as demonstrated by experiments on four datasets and ablation studies. Specifically, our method achieves absolute performance gains of 2.76% on MoNuSeg, 3.12% on DSB, 2.87% on Electron Microscopy, and 4.03% on TNBC datasets compared to existing SOTA methods.

Whole-body pose estimation is a challenging task that requires simultaneous prediction of keypoints for the body, hands, face, and feet. Whole-body pose estimation aims to predict fine-grained pose information for the human body, including the face, torso, hands, and feet, which plays an important role in the study of human-centric perception and generation and in various applications. In this work, we present RTMW (Real-Time Multi-person Whole-body pose estimation models), a series of high-performance models for 2D/3D whole-body pose estimation. We incorporate RTMPose model architecture with FPN and HEM (Hierarchical Encoding Module) to better capture pose information from different body parts with various scales. The model is trained with a rich collection of open-source human keypoint datasets with manually aligned annotations and further enhanced via a two-stage distillation strategy. RTMW demonstrates strong performance on multiple whole-body pose estimation benchmarks while maintaining high inference efficiency and deployment friendliness. We release three sizes: m/l/x, with RTMW-l achieving a 70.2 mAP on the COCO-Wholebody benchmark, making it the first open-source model to exceed 70 mAP on this benchmark. Meanwhile, we explored the performance of RTMW in the task of 3D whole-body pose estimation, conducting image-based monocular 3D whole-body pose estimation in a coordinate classification manner. We hope this work can benefit both academic research and industrial applications.

Driver drowsiness is a major factor that contributing to road accidents. Several researches are ongoing to detect driver drowsiness, but they suffer from the complexity and cost of the models. This paper introduces a hybrid artificial intelligence (AI)-driven framework integrating deep learning (DL) and machine learning (ML) models for real-time drowsiness detection. The system utilizes a robust DL model to classify driver states based on facial images and support vector machine (SVM) model is trained to develop a cost-efficient yet robust facial landmark detector to extract key features such as eye aspect ratio (EAR) and mouth aspect ratio (MAR). We also introduce a multi-stage decision fusion mechanism that combines convolutional neural network (CNN) probability scores with EAR/MAR thresholds to enhance detection reliability and reduce false positives. Experimental results demonstrate that the proposed model achieves 98% accuracy and F1-score, significantly outperforming traditional DL approaches. Additionally, the SVM-based landmark predictor shows improved efficiency with lower mean squared error (MSE) without having higher computational requirements.

This article presents a compression-based adaptive algorithm for Chinese Pinyin input. There are many different input methods for Chinese character text and the phonetic Pinyin input method is the one most commonly used. Compression by Partial Match (PPM) is an adaptive statistical modelling technique that is widely used in the field of text compression. Compression-based approaches are able to build models very efficiently and incrementally. Experiments show that adaptive compressionbased approach for Pinyin input outperforms modified Kneser-Ney smoothing method implemented by SRILM language tools .

A realistic analysis of media access control protocols for wireless applications must incorporate the affects of physical layer phenomena such as Rayleigh fading, shadowing and other processes that cause signal attenuation. In this work we present an innovative technique where the effects of the physical layer are embedded in the link layer by creating a fictitious user that blocks channel access from other users similar to the way fading inhibits real users from channel access. We match the fading process with the user's status, by minimizing the cross entropy prior to analyzing message delay. Simulations are used to assess the accuracy of the approach.

“Neutrosophic Sets and Systems” (NSS) Vol. 76 (2025) presents the latest scholarly contributions to neutrosophy and its interdisciplinary extensions. The issue showcases original research on neutrosophic set theory, neutrosophic logic, probability, and statistics, together with novel applications in data mining, artificial intelligence, and decision making under uncertainty. Papers explore fuzzy based correlation analysis, neutrosophic pattern mining, and weighted frequent itemset mining on uncertain databases, illustrating how neutrosophic frameworks can capture indeterminacy, neutrality, and contradiction beyond classical and intuitionistic fuzzy models. The volume also includes methodological advances, case studies, and theoretical discussions that deepen the mathematical foundations of neutrosophic systems and broaden their applicability across engineering, computer science, and the natural sciences. As an open access, SCOPUS indexed journal, NSS continues to foster rigorous, peer reviewed dissemination of neutrosophic research worldwide.

In this paper we analyze the performance of three algorithms for soft evidential update, in which a probability distribution represented by a Bayesian network is modified to a new distribution constrained by given marginals, and closest to the original distribution according to cross entropy. The first algorithm is a new and improved version of the big clique algorithm [1] that utilizes lazy propagation . The second and third algorithm [3] are wrapper methods that convert soft evidence to virtual evidence, in which the evidence for a variable consists of a likelihood ratio. Virtual evidential update is supported in existing Bayesian inference engines, such as Hugin. To evaluate the three algorithms, we implemented BRUSE (Bayesian Reasoning Using Soft Evidence), a new Bayesian inference engine, and instrumented it. The resulting statistics are presented and discussed.

Contrastive losses yield state-of-the-art performance for person re-identification, face verification and few shot learning. They have recently outperformed the cross-entropy loss on classification at the ImageNet scale and outperformed all self-supervision prior results by a large margin (SimCLR). Simple and effective regularization techniques such as label smoothing and self-distillation do not apply anymore, because they act on multinomial label distributions, adopted in cross-entropy losses, and not on tuple comparative terms, which characterize the contrastive losses. Here we propose a novel, simple and effective regularization technique, the Class Interference Regularization (CIR), which applies to cross-entropy losses but is especially effective on contrastive losses. CIR perturbs the output features by randomly moving them towards the average embeddings of the negative classes. To the best of our knowledge, CIR is the first regularization technique to act on the output features. In experimental evaluation, the combination of CIR and a plain Siamese-net with triplet loss yields best few-shot learning performance on the challenging tieredImageNet. CIR also improves the state-of-the-art technique in person re-identification on the Market-1501 dataset, based on triplet loss, and the state-of-the-art technique in person search on the CUHK-SYSU dataset, based on a cross-entropy loss. Finally, on the task of classification CIR performs on par with the popular label smoothing, as demonstrated for CIFAR-10 and -100.

Contrastive losses yield state-of-the-art performance for person re-identification, face verification and few shot learning. They have recently outperformed the cross-entropy loss on classification at the ImageNet scale and outperformed all self-supervision prior results by a large margin (SimCLR). Simple and effective regularization techniques such as label smoothing and self-distillation do not apply anymore, because they act on multinomial label distributions, adopted in cross-entropy losses, and not on tuple comparative terms, which characterize the contrastive losses. Here we propose a novel, simple and effective regularization technique, the Class Interference Regularization (CIR), which applies to cross-entropy losses but is especially effective on contrastive losses. CIR perturbs the output features by randomly moving them towards the average embeddings of the negative classes. To the best of our knowledge, CIR is the first regularization technique to act on the output featu...

This work proposes real time fixed yoga pose estimation and correction system. The pose estimation approaches, including MoveNet and OpenPose, are likely to suffer from low-quality keypoint localization when there are complex body poses or occlusions, and they trade off accuracy for real-time performance. In surpassing these drawbacks, this system improves keypoint prediction by incorporating anatomical priors into the heatmap generation process. The pipeline combines a lightweight MoveNet model for real-time inference with a stacked hourglass network trained offline on a yoga pose dataset. During training, hybrid heatmaps are generated using Kernel Density Estimation (KDE) along with PCA-aligned geometric masks (ellipse and stadium) to create more anatomically meaningful supervision signals. At inference time, MoveNet keypoints are transformed into hybrid heatmaps and compared against predictions using an Adaptive Squared Mean Loss function that emphasizes precision at keypoint peaks. Quantitative performance on the MPII dataset confirms that the proposed method obtains a [email protected] of 89.3%, mAP of 74.2%, and NME of 0.101-better than MoveNet (PCK 84.1%), HRNet-W32 (PCK 88.5%), and OpenPose (PCK 83.9%). The system also operates at 22 FPS, leading to real-time performance with improved accuracy. Ablation experiments confirm the efficacy of hybrid heatmaps and mask geometries since the proposed loss function learns 20 epochs ahead of Wing Loss and 30 ahead of MSE. The method demonstrates great promise for yoga training and fitness feedback application scenarios, where speed and accuracy are of equal worth

Information theory provides principled models for different machine learning problems, such as clustering, dimensionality reduction, classification, and many others. However, using the definitions of information theoretic quantities directly, posses a challenging estimation problem that often leads to strong simplifications, such as Gaussian models, or the use of plug in density estimators that restrict the kind of representations that we can use on the data. In this work, we adopt a data-driven perspective based on reproducing kernel Hilbert space methods that leads to successful application of information theoretic principles without resorting to estimation of the underlying probability distributions. The proposed methodology offers several advantages compared to other state of the art work such as entropic graphs because it can provide quantities that are more amenable for optimization (differentiability) as well as the representation flexibility that kernel methods provide. The work is divided into two main parts. First, we introduce an information theoretic objective function for unsupervised learning called the principle of relevant information. We employ an information theoretic reproducing kernel Hilbert space (RKHS) formulation, which can overcome some of the limitations of previous approaches based on Parzen's density estimation. Results are competitive with kernel-based feature extractors such as kernel PCA. Moreover, the proposed framework goes further on the relation between information theoretic learning, kernel methods and support vector algorithms.

Out-of-distribution (OOD) detection is a critical aspect of deploying robust and reliable machine learning models. This paper investigates the efficacy of self-supervised pretraining using Momentum Contrast (MoCo) for OOD detection. We present a comprehensive pipeline where a ResNet-18 model is pretrained on the CIFAR-10 dataset using MoCo, followed by fine-tuning for classification. The pretrained and fine-tuned model is then evaluated for its OOD detection capabilities on the Street View House Numbers (SVHN) dataset. We employ three distinct OOD detection metrics: Mahalanobis distance, energy scores, and a novel hybrid Energy-Gradient Score (EGS). Our findings demonstrate that energy-based methods, particularly the EGS, significantly outperform the Mahalanobis distance in distinguishing between in-distribution and out-of-distribution data. This suggests that self-supervised pretraining with MoCo can learn rich and generalizable features that are highly effective for OOD detection, and that the choice of OOD detection metric plays a crucial role in leveraging the full potential of these learned representations.

Skin cancer represents one of the most common malignancies globally, making early detection crucial for effective treatment and improved patient outcomes. While dermatologists typically rely on dermoscopy and clinical examinations for diagnosis, recent advances in artificial intelligence, specifically deep learning techniques like convolutional neural networks (CNNs), have shown significant promise in automating skin lesion classification [1,2]. Although CNN models trained on benchmark dermatological datasets such as HAM10000 and ISIC have demonstrated diagnostic accuracies comparable to expert dermatologists, their effectiveness declines when faced with evolving realworld data distributions, a phenomenon known as concept drift [3,4]. To address the limitations associated with static AI models, this paper proposes a novel multi-agent deep learning framework designed for continual learning and adaptive skin lesion diagnosis. The architecture begins with multiple agents trained on trusted expert-annotated datasets, each subsequently specialized by continuous fine-tuning using distinct streams of dermatological images sourced from teledermatology platforms and social media. These crowdsourced datasets capture emerging dermatological conditions, varied imaging technologies, and diverse patient demographics, providing valuable but noisy real-world data. Crucially, the system includes a centralized Supervisor Agent responsible for periodically evaluating the performance of each specialized agent. Once annotated, these validated cases enrich the training datasets, enabling agents to continually adapt to new clinical trends and maintain robust diagnostic accuracy over time. The proposed multi-agent architecture thus integrates continual learning, domain adaptation, and expert oversight, effectively addressing concept drift and advancing practical, scalable AI-driven diagnostic support in dermatology.

In this paper, we consider the local fractional decomposition method, variational iteration method, and differential transform method for analytic treatment of linear and nonlinear local fractional differential equations, homogeneous or nonhomogeneous. The operators are taken in the local fractional sense. Some examples are given to demonstrate the simplicity and the efficiency of the presented methods.