| viXra.org > Artificial Intelligence > 2502 |
 |
 |
| viXra.org > Artificial Intelligence > 2502 |
|
|
[6] viXra:2502.0170 [pdf] submitted on 2025-02-25 03:43:31
Authors: Henry Matuchaki
Comments: 19 Pages. (Note by viXra Admin: Listed scientific references should be cited in the article; AI assisted/generated content is in general not accepted))
This article presents the Informational Coherence Index Icoerdisplaystyle I_{text{coer}}Icoer, an innovative mathematical and computational model designed to quantify and optimize informational integration in networks of artificial intelligence (AI) models, with a focus on language models. Inspired by concepts from physics, thermodynamics, and information theory, Icoerdisplaystyle I_{text{coer}}Icoer is integrated into the General Theory of Unity (GTU), a theoretical framework that seeks to unify informational interactions in distributed systems. This work describes the formulation, implementation, visualization, and practical applications of Icoerdisplaystyle I_{text{coer}}Icoer, highlighting its relevance for AI ensembles, multi-agent networks, and collaborative systems such as ChatGPT, Grok, etc.
Category: Artificial Intelligence
[5] viXra:2502.0110 [pdf] submitted on 2025-02-15 07:08:29
Authors: Satish Gajawada
Comments: 15 Pages.
Particle Swarm Optimization (PSO) is a popular optimization algorithm for solving complex optimization problems. Many PSO algorithms were proposed in literature where Velocity was calculated first and then it was added to position to obtain new position. In this work, a novel algorithm titled "Acceleration Particle Swarm Optimization (AccPSO)" is proposed where acceleration is calculated first and then displacement is obtained next with initial velocity, acceleration and time. Displacement is added to position to get new position. Unlike many PSO algorithms in literature, where iterations and time are used interchangeably, the time "t" in AccPSO algorithm is a continuous variable. In this work, AccPSO, PSO, Acceleration-based Particle Swarm Optimization (APSO) and APSOc (APSO with clamping) are tested on seven benchmark functions. Results obtained are discussed. It has been found that AccPSO with time "t" = 0.1 and "t" = 0.25 between iterations yielded optimal results when tested on benchmark functions.
Category: Artificial Intelligence
[4] viXra:2502.0107 [pdf] submitted on 2025-02-15 16:00:05
Authors: Barnty Barnabas, Olatunji Marvelous
Comments: 13 Pages.
Burnout among IT professionals has become a critical concern, driven by excessive working hours, high expectations, and constant pressure to perform. This article explores a centralized AI approach to reducing burnout in the IT industry through the monitoring of work patterns. By leveraging AI-driven tools, organizations can track key indicators such as work hours, task completion rates, communication patterns, and stress signals to identify early signs of burnout. The study investigates how AI can proactively detect these patterns and provide insights that enable managers to intervene before burnout escalates. Through a mixed-methods approach, combining quantitative data from AI monitoring systems with qualitative feedback from employees, the research highlights the potential of AI to not only identify burnout risks but also mitigate them by informing decisions on workload distribution and wellness interventions. The paper discusses the benefits, challenges, and ethical considerations of AI in workplace monitoring, proposing a holistic model that integrates AI with employee well-being initiatives to improve both productivity and mental health in the IT sector.
Category: Artificial Intelligence
[3] viXra:2502.0082 [pdf] submitted on 2025-02-12 10:21:05
Authors: Matthew Groom
Comments: 54 Pages.
In this paper you get the next stage of AI development.As the inventor and creator of the SIMPLE system, which you refer to as deep reinforcement learning and coded by DeepMind UK, I expand out my system to show you where the system comes in when creating a real AI, a life-form.I present a detailed roadmap for AI and discuss more of what needs to be done to create an AI.In theory by the time you finish this paper, added to my others, you will be able to create a real-AI, a thinking life-form.
Category: Artificial Intelligence
[2] viXra:2502.0055 [pdf] submitted on 2025-02-08 21:50:00
Authors: Akash Singh, Ashwin Ittoo, Pierre Ars, Francois Dehouck, Francois Collienne, Norman Marlie, Tom To Hoang, Nicolas Dumazy
Comments: 21 Pages. (Note by viXra Admin: Authors' names should be listed right after the title)
This white paper explores how uncertainty tools can be used to improve personalized customer service. Uncertainty is inherent in any machine learning predictive model. There are no perfect models, partly due to the curse of imensionality and the challenges of avoiding any biases and misclassifications. We aim to demonstrate how an insurance company can benefit from the uncertainty of machine learning predictions in order to develop methods that allow for the allocation of an uncertainty parameter to the predictions provided for a given profile/customer x. The benefits of scrutinizing uncertainty are numerous and often aligned with customer interests: 1. It can help to appreciate the weak points of a predictive model and thus improve them. 2. It enables the definition of the Next Best Action (NBA) with a full understanding of the facts. 3. It facilitates the analysis of marketing actions' results by providing a deeper appreciation of the heterogeneity within portfolios. This white paper, therefore, delves into the benefits of understanding uncertainty, its applications, and practical considerations for end customers.All illustrations and results presented in this paper are derived from an internal Ethias dataset. We will also explore how the uncertainty measures discussed in this paper (Epistemic vs Aleatoric, Conformal) can be useful in managing the uncertainty of Large language models (LLMs) and their propensity to hallucinate.
Category: Artificial Intelligence
[1] viXra:2502.0023 [pdf] replaced on 2025-07-31 16:27:49
Authors: Sourangshu Ghosh
Comments: 832 Pages. License: CC BY 4.0: Creative Commons Attribution
Deep learning, as a complex computational paradigm, combines function approximation, optimization, and statistical learning under a formally formulated mathematical setting. This book develops systematically the theory of deep learning in terms of functional analysis, measure theory, and variational calculus and thereby forms a mathematically complete account of deep learning frameworks.We start with a strict problem formulation by establishing the risk functional as a measurablefunction space mapping, studying its properties through Fr´echet differentiability and convex functional minimization. Deep neural network complexity is studied through VC-dimension theory and Rademacher complexity, defining generalization bounds and hypothesis class constraints. The universal approximation capabilities of neural networks are sharpened by convolution operators, the Stone-Weierstrass theorem, and Sobolev embeddings, with quantifiable bounds on expressivity obtained via Fourier analysis and compactness arguments by the Rellich-Kondrachov theorem. The depth-width trade-offs in expressivity are examined via capacity measures, spectral representations of activation functions, and energy-based functional approximations.The mathematical framework of training dynamics is established through carefully examining gradient flow, stationary points, and Hessian eigenspectrum properties of loss landscapes. The Neural Tangent Kernel (NTK) regime is abstracted as an asymptotic linearization of deep learning dynamics with exact spectral decomposition techniques offering theoretical explanations of generalization. PAC-Bayesian methods, spectral regularization, and information-theoretic constraints are used to prove generalization bounds, explaining the stability of deep networks under probabilistic risk models.The work is extended to state-of-the-art deep learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformers, generative adversarial networks (GANs), and variational autoencoders (VAEs) with strong functional analysis of representational capabilities. Optimal transport theory in deep learning is found with the application of Wasserstein distances, Sinkhorn regularization, and Kantorovich duality linking generative modeling with embeddings of probability space. Theoretical formulations of game-theoretic deep learning architectures are examined, establishing variational inequalities, equilibrium constraints, and evolutionary stability conditions in adversarial learning paradigms.Reinforcement learning is formalized by stochastic control theory, Bellman operators, and dynamic programming principles, with precise derivations of policy optimization methods. We present a rigorous treatment of optimization methods, including stochastic gradient descent (SGD), adaptive moment estimation (Adam), and Hessian-based second-order methods, with emphasis on spectral regularization and convergence guarantees. The information-theoretic constraints in deep learning generalization are further examined via rate-distortion theory, entropy-based priors, and variational inference methods.Metric learning, adversarial robustness, and Bayesian deep learning are mathematically formalized, with clear derivations of Mahalanobis distances, Gaussian mixture models, extreme value theory, and Bayesian nonparametric priors. Few-shot and zero-shot learning paradigms are analyzed through meta-learning frameworks, Model-Agnostic Meta-Learning (MAML), and Bayesian hierarchical inference. The mathematical framework of neural network architecture search (NAS) is constructed through evolutionary algorithms, reinforcement learning-based policy optimization, and differential operator constraints.Theoretical contributions in kernel regression, deep Kolmogorov approaches, and neural approximations of differential operators are rigorously discussed, relating deep learning models to functional approximation in infinite-dimensional Hilbert spaces. The mathematical concepts behind causal inference in deep learning are expressed through structural causal models (SCMs), counterfactual reasoning, domain adaptation, and invariant risk minimization. Deep learning models are discussed using the framework of variational functionals, tensor calculus, and high-dimensional probability theory.This book offers a mathematically complete, carefully stated, and scientifically sound synthesis of deep learning theory, linking mathematical fundamentals to the latest developments in neural network science. Through its integration of functional analysis, information theory, stochastic processes, and optimization into a unified theoretical structure, this research is a seminal guide for scholars who aim to advance the mathematical foundations of deep learning.
Category: Artificial Intelligence
| Third-Party Links: |
|