| viXra.org > Artificial Intelligence > 2510 |
 |
 |
| viXra.org > Artificial Intelligence > 2510 |
|
|
[6] viXra:2510.0139 [pdf] submitted on 2025-10-28 13:54:18
Authors: Ritvik Chappidi, Aditya Jupally
Comments: 2 Pages.
Scaling laws describe how model performance improves with dataset size, model width, and compute. While such laws are well documented for large-scale language models, their behavior in small networks remains less understood. This paper presents a concise empirical study of loss scaling behavior in simple feedforward neural networks trained on synthetic regression tasks. Results show that even very small networks follow an approximate power-law relationship between dataset size and test loss, with a fitted exponent of about 0.076. These findings suggest that scaling regularities emerge even at small scales, implying that the underlying principles of efficiency and generalization extend beyond large-scale models.
Category: Artificial Intelligence
[5] viXra:2510.0079 [pdf] submitted on 2025-10-15 20:39:39
Authors: Jaba Tkemaladze
Comments: 23 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)
This article presents the Ze artificial life system, a novel bio-inspired architecture for predictive processing in infinite data streams under severe memory constraints. The system implements Bayesian probability updating through a mechanism of dynamic chronotropic frequency analysis, demonstrating remarkable computational efficiency and biological plausibility. Unlike traditional approaches such as LSTM networks and Markov models, Ze processes information through parallel beginning and inverse processors, enabling complementary pattern discovery while maintaining sublinear memory complexity. The core algorithm exhibits distinctive probability dynamics characterized by an initial match probability of 0.5 with exponential decay to 0.00001 as counter diversity increases, achieving 78-92% prediction accuracy for stable data flows. Experimental results using synthetic datasets (1,048,576 binary sequences) confirm 37-42% operational savings compared to conventional methods, rapid adaptation to changing stream characteristics within 2-3 seconds, and robust noise tolerance up to 15% input distortion. The Go implementation processes 1.2 million operations per second with 850 nanosecond latency while maintaining memory usage of 12.8 bytes per counter. The system's architecture shows strong neurobiological correlations with predictive coding principles and synaptic plasticity mechanisms, providing both a practical solution for resource-constrained environments and a computational model of Bayesian inference in neural systems. Future development pathways include extension to non-binary data streams, integration with hierarchical Bayesian models, and hardware acceleration through memristor-based implementations.
Category: Artificial Intelligence
[4] viXra:2510.0049 [pdf] submitted on 2025-10-09 20:52:23
Authors: Ekam Chatterjee
Comments: 19 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)
Epileptic seizure detection from electroencephalogram (EEG) signals represents a fundamental challenge in computational neuroscience, with traditional approaches limited by their inability to capture complex topological transformations in brain connectivity during ictal events. While topological data analysis has demonstrated promise for EEG analysis, existing methodologies primarily employ persistent homology features with conventional classifiers, failing to leverage the geometric structure inherent in neuralcomputation. To the best of our knowledge, this is the first work that applies topological neural networks—message passing architectures on simplicial complexes—to EEG seizure detection, integrating persistent homology features across multiple distance functions with temporal modeling, building upon Hajij et al.’s foundational work on topological deep learning architectures. The proposed approach introduces a novel 3-layer TNN framework that integrates multi-scale persistent homology with theoretically grounded topological message passing mechanisms. This research establishes mathematical foundationsfor seizure detection through topological invariants and provides convergence guarantees for the neural architecture. The model constructs four complementary distance matrices (correlation, Euclidean, phaselag, and coherence-based) from multi-channel EEG recordings, applying Vietoris-Rips filtrations to extract multi-dimensional topological features across scales. The core innovation lies in the rigorous implementation of the four-step topological message passing framework: message computation, within-neighborhoodaggregation, between-neighborhood aggregation, and feature update, combined with bidirectional LSTMnetworks for temporal modeling. Evaluation on the CHB-MIT dataset across 10 patients using event-based metrics demonstrates an F1-score of 74.36%, establishing the first successful integration of topological neural architectures with neurological signal processing. Theoretical analysis reveals that seizure events exhibit characteristic changes in topological entropy and Betti numbers, providing interpretable biomarkers for clinical translation.
Category: Artificial Intelligence
[3] viXra:2510.0042 [pdf] submitted on 2025-10-08 06:08:54
Authors: Yilin Li
Comments: 10 Pages.
Mathematical and logical reasoning is an important component of human intelligence. Thus, a common metric for evaluating Large Language Models (LLMs) is their ability to solve mathematical problems. Recently, LLMs have shown remarkable performance in completing various tasks such as text generation, text understanding and image analysis. Their mathematical and reasoning ability has also advanced rapidly, allowing them to solve complex algebra problems. However, LLMs still exhibit limitations in describing and reasoning about geometric and spatial concepts, failing to accurately identify and understand the logic within geometric figures. In order to address this gap in understanding, numerous diverse datasets of geometric figures and metadata are needed to continue training their geometric reasoning capabilities. In this research paper, we introduce an innovative algorithm to create synthetic polygon geometric shape datasets, and define methods to integrate synthetic geometric images and metadata into major LLMs for training, validation, and evaluation of their geometric reasoning abilities.
Category: Artificial Intelligence
[2] viXra:2510.0040 [pdf] submitted on 2025-10-08 18:31:01
Authors: Geraldine Geoffroy
Comments: 12 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)
This paper proposes a novel architecture for distributed, traceable, and event-driven execution of LLM-related tasks by combining W3C Linked Data Notifications (LDN) with remote Python scripts executed via uv run. This architecture enables any AI task-especially inference- to be executed locally with no software installation, traced via interoperable notifications, and archived with full provenance metadata (e.g., models, parameters, etc.). To achieve this, the system leverages LDN as a semantic pub-sub orchestration layer, combined with uv-based scripts as reproducible, stateless microservices. We demonstrate the value of this architecture for building transparent, auditable, and distributed Large Language Model (LLM) inference workflows with three working proof-of-concepts: (1) a basic semantically-notified inference where notifications populate a register of evidences for transparency, (2) a Retrieval-Augmented Generation (RAG) pipeline triggered by Create events and executed through script-based stages, and (3) a distributed inference setup where task-specific SLMagents independently process jobs and respond via Announce messages. Each stage archive full provenance metadata (model version, script SHA, parameters, runtime) using PROV-O, supporting reproducibility and auditability. This architecture lays the groundwork for a lightweight, decentralized, and FAIR-aligned standard for orchestrating LLM tasks.
Category: Artificial Intelligence
[1] viXra:2510.0039 [pdf] submitted on 2025-10-08 18:28:36
Authors: Aldrich K. Wooden Sr.
Comments: 4 Pages.
Abstract—The convergence of rural healthcare andenvironmental monitoring demands an integrated,AI—native architecture spanning Critical Access Hospitals (CAHs) and rural water utilities. Despite near—universal basic EHR adoption in hospitals, only a minority of CAHs fully exchange data, while a large share of rural water utilities face critical cybersecuritydeficiencies and many rural areas lack minimum broadband capacity required for modern operations [R1—R3]. This white paper synthesizes technical requirements and a reference architecture (ARIS—2025 ) across connectivity, edge computing, data interoperability, andcompliance, mapping vendor ecosystems, cost benchmarks, and phased implementation to achieve resilient, privacy—preserving cross—sector analytics.
Category: Artificial Intelligence
| Third-Party Links: |
|