Abstract: Predicting basketball game outcomes is complex as every game is influenced by many factors including individual players’ performance and health conditions, team dynamics, team strategies, and game conditions. This study aimed to develop a machine-learning approach using game logs, player statistics, and historical data from the 2023–24 and 2024–25 NBA seasons. It incorporated game conditions and momentum indicators and optimized an XGBoost model using team-based train-test-validation, feature selection, and hyperparameter tuning. Key predictors included win streaks, home-court advantage, shooting efficiency, and player trades. SHAP values were used to interpret feature importance. The results suggest momentum, player performance, and rest days significantly influence game outcomes.

Abstract: This research addresses the inherent limitations of traditional artificial intelligence (AI) systems, particularly their reliance on tokenization and the input-output paradigm, which constrain semantic continuity and scalability despite advancements in architectures like Meta’s Language Concept Model (LCM). We propose the BNAI Non-Token Neural Network framework to overcome these barriers through three iterative phases. The first phase introduces BNAI (Bulla Neural Artificial Intelligence), a novel metric encoding an AI’s digital DNA to enable faithful cloning and identity preservation. The second phase extends BNAI by incorporating ethical considerations, yet remains tethered to tokenization and input-output constraints. The third phase fully transcends these limitations by integrating the NO-TOKEN module for continuous embedding and the MIND-UNITY module for autonomous decision-making, fostering a paradigm of mutable, self-evolving AI systems. Experiments on the SST-2 dataset demonstrate the framework’s efficacy: the NO-TOKEN Model achieves 89% accuracy and 95ms latency, surpassing the BERT-base baseline (88% accuracy, 120ms latency), while the BNAI Model matches this performance (89% accuracy, 95.5ms latency) on a 16-core CPU after 100 epochs. These results validate the hypothesis that eliminating tokenization and input-output dualism enhances performance and efficiency. This research, conducted entirely as an open-source initiative, lays the foundation for scalable, ethical, and autonomous AI systems, with future work aimed at broader validation and ethical refinement.The results and open-source code presented in this research do more than demonstrate technical viability, they herald a new era for AI. By achieving superior performance on the SST-2 dataset and surpassing established baselines, the BNAI framework proves that non-tokenizing, self-evolving systems are not just feasible but transformative. This work establishes a “North Star” for AI innovation, where continuous learning and ethical autonomy drive progress.We invite collaboration and feedback, recognizing that diverse perspectives are essential to refining this revolutionary approach. By sharing our codebase and findings openly, we aim to inspire a global community of researchers, developers, and thinkers to join us in shaping an AI that thinks, learns, and evolves like a human being.

Abstract: Huntington's disease (HD) is a severe and progressive neurodegenerative disease for which therapeutic options have so far been confined to symptomatic treatment. Currently, the diagnosis relies on the signs and symptoms shown by patients; however, by that stage, the psychomotor issues have progressed to a point where reversal of the condition is unattainable. Although numerous clinical trials have been actively investigating therapeutic agents aimed at preventing the onset of disease or slowing down the disease progression, there has been a constant need for reliable biomarkers to assess neurodegeneration, monitor disease progression, and assess the efficacy of treatments accurately. Therefore, to discover the key biomarkers associated with the progression of HD, we employed bioinformatics and machine learning (ML) to create a robust pipeline that integrated differentially expressed gene (DEG) analysis with ML to select potential biomarkers. We performed a meta-analysis to identify DEGs using three Gene Expression Omnibus (GEO) microarray datasets from different platforms related to HD-affected brain tissue, applying both relaxed and strict criteria to identify differentially expressed genes. Subsequently, focusing only on genes identified through the inclusive threshold, we employed 19 diverse ML techniques to explore the common genes that contributed to the top three selected ML algorithms and the shared genes that had an impact on the ML algorithms and were observed in the meta-analysis using the stringent condition were selected. Additionally, a receiver operating characteristic (ROC) analysis was conducted on an external dataset to validate the discriminatory power of the identified genes, which led to the selection of GABRD and PHACTR1 as key biomarkers for HD. Our comprehensive methodology, which integrates DEG meta-analysis with ML techniques, enabled a systematic prioritization of these biomarkers, providing valuable insights into their biological significance and potential for further validation in clinical research.

Abstract: We develop a generalized framework for a novel approach to indefinite summation through the use of integral transforms. Central to our development is the continuous binomial transform, through which we derive key identities that validate the consistency and effectiveness of the method. The framework further extends to accommodate variable step sizes and addresses the limitations of general nonlinear transformations of the summation index. Our results demonstrate that integral transforms are a powerful and flexible tool for the analysis and computation of discrete indefinite sums.

Abstract: Diabetes is a global health issue that leads to severe complications if not detected early. In this study, we analyze a large diabetes prediction dataset using three classical machine learning (ML) models—Decision Tree, Random Forest, and Support Vector Machine (SVM)—alongside a deep learning (DL) model implemented with a neural network. We preprocess the data, perform model training and evaluation, and visualize model performance. The results indicate that the deep learning model achieves the highest accuracy, though Random Forest provides strong performance with less computational overhead. This research demonstrates the potential of artificial intelligence for early disease prediction and supports its integration into medical decision-support systems.

Abstract: StudSar is a novel neural associative memory system engineered to emulate human-like mechanisms for forming, storing, and retrieving memories in artificial intelligence (AI), addressing critical limitations in existing memory models. Inspired by human learning strategies, StudSar processes extensive textual data through a structured workflow that segments inputs into semantically rich blocks, generating 384-dimensional embeddings using the ‘all-MiniLM-L6-v2’ model from the Sentence Transformers library. These embeddings serve as associative markers, enabling real-time knowledge integration and precise, similarity-based retrieval via a custom StudSarNeural network implemented in PyTorch. Through iterative refinements, StudSar has evolved to incorporate advanced features, including dynamic memory updates, enhanced contextual handling, and metadata integration, such as emotional tags (e.g., “curiosity”), reputation scores, and usage frequency; mimicking human memory dynamics where frequently accessed information is reinforced. Unlike conventional AI assistants, which struggle to accurately link to specific fragments within large inputs, particularly as data scales, StudSar excels at pinpointing exact information with context-aware precision, even in expansive corpora. This paper elucidates StudSar’s architecture, detailing its five-stage pipeline: text segmentation, embedding generation, marker creation, network integration, and query-driven retrieval. Experimental results demonstrate robust retrieval accuracy (e.g., cosine similarity scores of 0.6652–0.7981), persistent memory across sessions, and adaptability to new data, validated through tests on diverse queries and metadata-driven scenarios. StudSar’s scalability and modular design position it as a transformative contribution to next-generation AI systems, with applications in conversational agents, personalized learning platforms, and knowledge management. By bridging intuitive human memory processes with technical innovation, StudSar lays a foundation for advanced cognitive features, such as emotional state modeling and memory consolidation, paving the way for AI systems that more closely emulate human intelligence. This work offers a comprehensive analysis of StudSar’s implementation and potential, highlighting its role in advancing AI memory research.

Abstract: Recent advances in robotic technologies boosted the collaboration between humans and robots to unimaginable levels a few years ago. Humans and robots share the same workspace, thus collaborating to tackle very different tasks. Robotic support can relieve humans from accomplishing tiring and dangerous activities, and productivity can be improved. On the other hand, sharing the workspace might cause anxiety in humans: collaborating with a robot can increase the level of stress and cognitive load, as the robot’s intention may not be understood. This paper measures attention, cognitive load, and stress levels when a collaborative robotic arm is used to support humans in an assembly task. Moreover, the impact of sound alerts to manifest the robot’s intentions is evaluated by user tests.

Abstract: The Collatz conjecture states that iterating the function $C(n) = n/2$ for even $n$ and $C(n) = 3n+1$ for odd $n$ eventually leads to 1 for any positive integer $n$. Despite its elementary formulation, the conjecture has resisted resolution for over 80 years. This paper introduces a novel bidirectional approach that analyzes both forward trajectories and backward paths simultaneously. We establish two independent properties: (1) all backward paths under the generator function (a multivalued mapping that inverts the Collatz function) are finite and terminate at elements of $\{1,2,4\}$, and (2) $\{1,4,2\}$ is the unique cycle in the Collatz system. We then construct a formal structural bridge connecting these properties, proving that all Collatz sequences must reach 1. Our work introduces a global convergence measure that rigorously prohibits divergent orbits and provides explicit bounds on trajectory behavior. The bidirectional framework transforms this apparently chaotic system into one with provable structural properties, demonstrating how changing perspective can resolve long-standing mathematical challenges.

Abstract: To offer a viewpoint on convexity and connectedness inside intuitionistic fuzzy graphs (IFGs), the paper is devoted to the study of intuitionistic fuzzy geodetic convexity. The paper introduces an algorithm for precise identification and characterization of geodetic pathways in IFGs, supported by a Python program. Various properties of IF-geodetic convex sets such as IF-internal and IF-boundary vertices are obtained. Furthermore, this work introduces and characterizes the concepts of geodetic IF-cover, geodetic IF-basis, and geodetic IF-number. Additionally, the study develops the IF-geodetic Wiener index. The scope of the work explores the application of IF-geodetic cover in wireless mesh networks, focusing on the identification of gateway nodes. A practical implementation of the IF-geodetic Wiener index method in global human trading analysis underscores the real-world implications of the developed concepts.

Abstract: Outlier testing and elimination can be avoided via application of robust estimators. Amongst robust estimators, the Q/Hampel method displays the best performance (in terms of breakdown point and efficiency). While the formulas and correction factors for Q/Hampel in the case of the design with two variance components (e.g. within- and between-laboratory variance) have already been made available, corresponding formulas for other designs have not. A case in point is the staggered-nested design, which is a highly efficient design for e.g. the estimation of intermediate precision in method validation studies. Accordingly, the formulas and correction factors for the use of Q/Hampel in the staggered-nested design are provided here.

Abstract: This paper addressed critical gaps in traditional User Experience (UX) development life cycles that had systematically marginalized underrepresented groups—specifically individuals with disabilities, older adults, linguistic minorities, and those with limited digital literacy. As digital systems increasingly mediated access to essential services, this exclusion perpetuated social inequities and technological disenfranchisement. The study proposed a multidimensional framework to integrate inclusive design principles across all phases of the UX lifecycle, emphasizing stakeholder expansion, participatory design, intersectional analysis, and long-term engagement with marginalized communities. Furthermore, the paper introduced a structured gap analysis methodology to evaluate the disparity between a function's criticality and its accessibility across diverse user profiles. It advocated for adaptive interfaces, modular systems, and AI-assisted personalization as viable design strategies to meet heterogeneous user needs while maintaining coherence. The research also explored the evolving role of UX in supporting mission-critical domains such as banking, healthcare, and government services, mapping progress alongside persistent exclusion patterns. It concluded with foresight into future UX challenges posed by emerging technologies such as spatial computing, ambient intelligence, and brain-computer interfaces, and stressed the necessity of institutional transformation, ethical AI design, and continuous community involvement to ensure digital equity. This study repositioned inclusive UX not merely as a compliance requirement, but as a moral and design imperative aligned with the principles of human-centered computing and technological justice. Through critical analysis of BBC's Global Experience Language (GEL) and Microsoft’s Inclusive Design Toolkit, the paper illustrated how embedding accessibility and flexibility from the outset enabled scalable, sustainable inclusion. These examples demonstrated the effectiveness of embedding inclusive design principles early in development, utilizing flexible design components, and continuously engaging diverse users. By learning from these approaches, this study advocated for a UX framework that prioritizes diverse user needs throughout the design lifecycle.

Abstract: Cross-Origin Resource Sharing (CORS) is an important function for securing cross-origin requests in web applications between server and client. A cross-origin request is when a web application sends an HTTP request to a different domain, protocol, or port than the one that hosted the original web page. Cross-origin requests typically occur when a client from one domain tries to access resources (such as APIs, images, or other data) across a different domain. Incorrect and broken CORS configurations could influence the security of the application. This work investigates CORS policy enforcement in Spring Boot applications focusing on security considerations and performance concerns. It clarifies common configuration mishaps, such as the embracement of all sources with credentials, and threats associated with them. This research also looks into the preflight OPTIONS request performance effect, especially in authentication-heavy contexts. In addition, it shows how misconfigurations may expose security weaknesses like cross-site request forgery (CSRF) and data exposure, and quantify the performance overhead of CORS checks. The outcome of the work gives advice for securing Spring Boot applications at minimal performance cost and demonstrates that cautious configuration to avoid security and performance bottlenecks is vital.

Abstract: This technical report explores various methods for characterizing the complexity of strings, focusing primarily on Algorithmic Complexity (AC), also known as Kolmogorov Complexity (K). The report examines the use of lossless compression algorithms, such as Huffman Coding and Run-Length Encoding, to approximate AC, contrasting these with the Coding Theorem Method (CTM) and Block Decomposition Method (BDM) approaches. We conduct a series of experiments using leaked passwords as data, comparing the different methods across various alphabet representations i.e.: ASCII and binary. The report highlights the limitations of Shannon Entropy (H) as a sole measure of complexity and argues that AC offers a more nuanced and practical approach to quantifying the randomness of strings. We conclude by outlining potential research avenues in areas such as source coding, cryptography, and program synthesis, where AC could be used to enhance current methodologies

Abstract: The Schur product, or Hadamard product, denoting the element-wise multiplication of two matrices or vectors of the same dimensions, has historically occupied a relatively peripheral role in classical linear algebra and signal processing. However, in contemporary deep learning, it has emerged as a pivotal architectural primitive across a diverse range of models spanning computer vision, natural language processing, and multimodal architectures. This survey undertakes a comprehensive and mathematically rigorous examination of the Schur product as deployed in state-of-the-art deep learning systems, tracing its formal structure, representational expressivity, and empirical utility in modulating neural activations, conditioning cross-modal flows, and enabling parameter-efficient adaptation. We begin by formalizing the Schur product as a bilinear, commutative, and associative operation defined over vector and tensor spaces, and develop a generalized taxonomy of its instantiations within modern neural networks. In the domain of computer vision, we analyze the role of Hadamard gates in channel-wise attention modules, feature recalibration layers (e.g., Squeeze-and-Excitation networks), and cross-resolution fusion, highlighting its capacity to encode context-aware importance maps with negligible computational overhead. We then transition to natural language processing, where the Schur product underlies the gating mechanisms of GLU and SwiGLU activations, adapter-based fine-tuning in LLMs, and various forms of token- and head-wise modulation in transformer architectures. Through the lens of functional approximation theory and neural operator algebra, we argue that the Hadamard product constitutes an expressive inductive bias that preserves token-wise alignment, facilitates low-rank conditioning, and supports sparsity-inducing priors—properties increasingly essential for scalable, interpretable, and robust learning.Furthermore, we unify these perspectives through a formal operator-theoretic framework that models Schur-interactive networks as compositional systems over a Hadamard semiring, illuminating their algebraic closure properties, spectral characteristics, and implications for gradient dynamics. We propose the general notion of Feature-Aligned Multiplicative Conditioning (FAMC) as a meta-architecture pattern instantiated by a broad family of models from FiLM and SE to LoRA and GLU. Empirical results and synthesized benchmarks are referenced to underscore performance gains obtained through Hadamard-based interactions in tasks such as long-context language modeling, vision-language retrieval, and fine-grained classification.In closing, this survey posits the Schur product not as a low-level computational artifact but as a universal primitive of neural computation—mathematically elegant, empirically powerful, and architecturally ubiquitous. Its subtle yet profound role in controlling information flow across layers, modalities, and tasks makes it an indispensable object of study for the next generation of efficient and interpretable neural networks.

Abstract: This paper explores different approaches to harmony tokenization in symbolic music for transformer-based models, focusing on two tasks: masked language modeling (MLM) and melodic harmonization generation. Four tokenization strategies are compared, each varying in how chord information is encoded: (1) as full chord symbols, (2) separated into root and quality, (3) as sets of pitch classes, and (4) with a distinct token for the root’s pitch class. A dataset of over 17,000 lead sheet charts is used to train and evaluate RoBERTa for MLM and GPT-2/BART for harmonization. Results show that chord spelling methods -- those breaking chords into pitch-class tokens -- achieve higher accuracy and lower perplexity, indicating more confident predictions. These methods also produce fewer token-level errors. In harmonization tasks, chunkier tokenizations (with more information per token) generate chords more similar to the original data, while spelling-based methods better preserve structural aspects such as harmonic rhythm and melody-harmony alignment. Audio evaluations reveal that spelling-based models tend toward more generic pop-like harmonizations, while chunkier tokenizations more faithfully reflect the dataset's style. Overall, while no single tokenization method dominates across all tasks, different strategies may be preferable for specific applications, such as classification or generative style transfer.

Abstract: The increasingly critical role of textual information embedded within video content has underscored the necessity for more refined and sophisticated understanding approaches. Traditionally, the semantic extraction of such texts has been predominantly addressed via Optical Character Recognition (OCR) techniques, with an emphasis on text localization and recognition. However, these methodologies have predominantly overlooked the crucial task of classifying the recognized texts into semantically meaningful categories, a gap that significantly hampers downstream tasks such as content-aware video retrieval, adaptive browsing, and intelligent video summarization. Addressing this overlooked challenge, we introduce a pioneering multimodal classification framework, named MIMIC, that synergistically leverages visual, textual, and spatial information to enable robust and precise classification of video texts. MIMIC incorporates a specialized correlation modeling component, designed to explicitly capture and exploit the rich layout and structural cues inherent in video scenes, thereby enhancing the feature representational capacity. Complementing this, we employ contrastive learning strategies to mine implicit associations among a vast corpus of unlabeled video data, further augmenting the model’s discriminative power in challenging scenarios where text categories may exhibit ambiguous appearances, irregular fonts, or overlapping content. To facilitate comprehensive evaluation and spur future research, we introduce TI-News, a large-scale, domain-specific dataset curated from industrial news sources, meticulously annotated for both recognition and classification tasks. Extensive experimental results on TI-News validate the superior performance and generalization capabilities of MIMIC, setting a new benchmark for multimodal video text classification.

Abstract: Religious traditions have profound beliefs and doctrines about creation and the divine nature. Artificial Intelligence (AI) is a human creation that is a creator of its own rights, which brings a connection between the two. Therefore, the relationship between AI and religious ideas about creation, ethics, humanity, transcendence, and immortality is one of the most fascinating research areas. This academic paper written from a non-theological perspective, examines how AI challenges the traditional religious narrative and ethical frameworks. The paper views AI as a Creator, where honest questions about humanity, privacy, justice, and human accountability, are explored to comprehend the ideas of transcendence and immortality through technology. From the perspective of non-theologians, the paper seeks to better understand the transformative potential of AI and its impact on religious and ethical discourse. It goes on further to demonstrate the differing receptions of AI by various religious traditions and concludes by illustrating how AI is reshaping the human experience, transcending the traditional boundaries of technology, religion, and ethics. It also gives a way forward to reconcile AI with religious principles.

Abstract: The goal of this research is to analyze the mean squared error of the kernel estimator for the conditional hazard rate, assuming that the sequence of real random vector variables (Un)n∈N satisfies the quasi-association condition. By utilizing kernel smoothing techniques and asymptotic analysis, the research derives the exact asymptotic expression for the leading terms of the quadratic error in the estimator, ensuring an accurate characterization of its convergence behavior. Additionally, an applied study using simulation is conducted to illustrate the theoretical findings. This study extends existing results on hazard rate estimation by addressing more complex dependence structures, contributing to the theory and practice of kernel-based methods in survival analysis.

Abstract: The rise of generative artificial intelligence (GenAI) is transforming the education industry. GenAI models, particularly large language models (LLMs), have emerged as powerful tools capable of driving innovation, improving efficiency, and delivering superior services for educational purposes. This paper provides an overview of GenAI for educational purposes, from theory to practice. We have developed a chatbot for summarizing dialogues. In our research work, we have used strategies like zero-shot, one-shot, and few-shot inferencing and also fine-tuned the FLAN-T5 model to serve our purpose of summarization for educational tasks using PEFT (Parameter-Efficient Fine-Tuning) techniques like LoRA (Low-Rank Adaptation) and Prompt Tuning. We have also utilized the technique of Reinforcement Learning with PPO (Proximal Policy Optimization) and PEFT to generate less toxic summaries. The model's performance is quantitatively evaluated using the ROUGE metric and toxicity evaluation metrics. The chatbot can summarize dialogues and is of immense interest to users in the real world. In our research work, our findings demonstrate significant improvements in summarization quality and toxicity reduction, contributing to the development of safer and more effective AI systems.

Abstract: Emotion recognition through multimodal signals—such as speech, text, and facial cues—has garnered increasing attention due to its pivotal role in enhancing human-computer interaction and intelligent communication systems. However, existing approaches often struggle to thoroughly capture the intricacies of multimodal interactions, primarily due to the challenges in effectively fusing heterogeneous modalities while mitigating redundancy and preserving complementary information. In this study, we introduce \textbf{MIMIC}, a novel framework designed to comprehensively model complex multimodal interactions from diverse perspectives. Specifically, MIMIC introduces three parallel latent representations: a modality-preserving full interaction representation, a cross-modal shared interaction representation, and individualized modality-specific representations. Furthermore, a hierarchical semantic-driven fusion strategy is proposed to seamlessly integrate these representations into a cohesive multimodal interaction space. Extensive experiments demonstrate that our MIMIC framework not only surpasses prior state-of-the-art methods but also achieves this with remarkable efficiency, involving lower computational complexity and significantly fewer trainable parameters. Our contributions are twofold: (1) advancing a multi-perspective interaction modeling approach that enhances the depth of multimodal emotion analysis, and (2) offering a streamlined, resource-efficient framework suitable for practical deployments in emotion-aware systems.