| viXra.org > Artificial Intelligence > 2409 |
 |
 |
| viXra.org > Artificial Intelligence > 2409 |
|
|
[11] viXra:2409.0161 [pdf] submitted on 2024-09-29 00:14:02
Authors: Hamidreza Seiti, Reza Javadi, Hossein Ghanbari, Sina Keshavarz
Comments: 56 Pages. In Chinese (Converted to pdf by viXra admin - Please submit article in pdf format only)
Supply chain risk management is a critical challenge in today’s increasingly complex and interconnected global markets, particularly within specific supply chains where disruptions can have far-reaching consequences. Generative Artificial Intelligence (GAI) transformer models have emerged as powerful tools for effectively managing these risks. However, selecting the most suitable GAI model for specific supply chain contexts remains a significant challenge due to the diverse range of available models and the complex interplay of risk factors involved. This challenge is further compounded by the necessity of considering human-centric criteria to ensure that the chosen model aligns with ethical standards and practical needs. This paper addresses this challenge by introducing an enhanced multi-criteria decision-making (MCDM) framework that refines the Evaluation based on Distance from Average Solution (EDAS) method. Our approach first improves the logical structure of the EDAS method and then incorporates the interactions and interdependencies between criteria, thereby overcoming key limitations of traditional MCDM methods and providing a more accurate and comprehensive evaluation process. We applied this improved EDAS model to the task of selecting the best GAI transformer model for risk management in the food supply chain. Through a systematic evaluation of various GAI models, considering their performance across multiple risk factors, our study identified GPT (Generative Pre-trained Transformer) as the most suitable model for this context, demonstrating superior capabilities in addressing the complex challenges associated with food supply chain risks. This research not only advances the theoretical foundation of MCDM techniques but also offers practical insights into the application of AI in supply chain management, highlighting the importance of human-centric AI approaches that prioritize transparency, ethical alignment, and effective decision-making.
Category: Artificial Intelligence
[10] viXra:2409.0158 [pdf] submitted on 2024-09-28 20:16:18
Authors: Meir Dudai
Comments: 46 Pages.
This paper explores the transformative potential of AI-powered underwriting engines in revolutionizing credit decisioning processes for embedded lending. Traditional methods of credit assessment often fall short in accurately evaluating creditworthiness, particularly for underserved populations. AI-powered underwriting engines address these limitations by leveraging machine learning algorithms and alternative data sources to provide more comprehensive and nuanced credit evaluations. This study examines the current landscape of credit decisioning, identifying key challenges and presenting a detailed analysis of AI-powered underwriting engines, including their technical architecture, key features, and potential for improving accuracy, speed, and inclusivity in lending decisions. The paper also considers implementation strategies, potential business impacts, and critical risk and compliance considerations. Finally, it looks ahead to future directions and scalability of AI-powered underwriting engines, considering emerging technologies and evolving regulatory landscapes.Index Terms—AI, credit decisioning, embedded lending, financial inclusion, machine learning, underwriting engines
Category: Artificial Intelligence
[9] viXra:2409.0107 [pdf] submitted on 2024-09-20 04:44:00
Authors: Reza Safdari, Mohammad Koohi-Moghaddam, Kyongtae Tyler Bae
Comments: 7 Pages.
In this study, we implemented a two-stage deep learning-based approach to segmentlesions in PET/CT images for the AutoPET III challenge. The first stage utilized aDynUNet model for coarse segmentation, identifying broad regions of interest. Thesecond stage refined this segmentation using an ensemble of SwinUNETR, SegResNet,and UNet models. Preprocessing involved resampling images to a common resolution andnormalization, while data augmentation techniques such as affine transformations andintensity adjustments were applied to enhance model generalization. The dataset was splitinto 80% training and 20% validation, excluding healthy cases. This method leveragesmulti-stage segmentation and model ensembling to achieve precise lesion segmentation,aiming to improve robustness and overall performance.
Category: Artificial Intelligence
[8] viXra:2409.0094 [pdf] submitted on 2024-09-17 08:58:05
Authors: Aryaman Sharma
Comments: 49 Pages.
Graph Neural Networks (GNNs) and reinforcement learning techniques are combined in GRAPPLE (GraphSAGE Reinforced with Actor-Proximal Policy Optimization), a revolutionary framework for improving personalized recommendation systems. GRAPPLE allows for dynamic adaptation to changing user preferences and item dynamics by fusing Proximal Policy Optimization (PPO) with GraphSAGE, a powerful representation learning technique. GRAPPLE can now efficiently extract extensive relational information from interaction graphs and capture complex user-item relationships. Adaptive learning techniques allow model to continuously update their recommendation criteria in response to user feedback, increase the precision of recommendations while maintaining the user satisfaction quota that it has. Experiments performed on real-world dataset demonstrate that our algorithm outperforms conventional recommendation methods, demonstrating its superiority in a range of recommendation scenarios as well as its durability and scalability. GRAPPLE represents a significant advancement in recommendation systems by combining GNNs with reinforcement learning methods. It provides a versatile and efficient way to manage interaction patterns and fluctuating user preferences in recommendation jobs.
Category: Artificial Intelligence
[7] viXra:2409.0086 [pdf] submitted on 2024-09-16 09:57:14
Authors: Mezbah Uddin Rafi
Comments: 16 Pages.
This paper examines the innovative application of Artificial Intelligence (AI) to simulate real-time historical what-if scenarios, exploring its potential for creating immersive and engaging educational experiences. AI-driven simulations could revolutionize the way history is taught, allowing users to engage directly with alternative historical outcomes. By exploring possible scenarios—such as different outcomes for major events like World War II or the Cuban Missile Crisis—students and educators can gain deeper insights into historical processes. This paper discusses the methodologies behind AI-driven historical simulations, the technical and ethical challenges involved, and the future potential of this technology.
Category: Artificial Intelligence
[6] viXra:2409.0073 [pdf] submitted on 2024-09-13 21:11:56
Authors: Sofiane Delloue
Comments: 20 Pages. (Author name added to the article by viXra Admin as required)
We introduce Newcoin, a novel protocol designed to accelerate open-source AI advancement by enabling the pooling of learning instances across diverse pipelines. This approach has the potential to multiply epistemic affordances exponentially, fostering unprecedented growth in AI capabilities. Newcoin leverages cryptographically signed statements and a game-theoretical consensus mechanism, which aggregates weighted human feedback to evaluate and reward network contributions. The open interpretability of learning signals contributes to improved generalization capabilities through several mechanisms. This shared cognitive space, where learning signals from various domains and tasks are universally interpretable, allows AI systems to leverage collective knowledge to better generalize to new, unseen problems. By integrating robust security measures with an incentive structure that promotes high-quality outputs, Newcoin creates a self-improving ecosystem for AI development. This innovative framework not only accelerates open-source AI capabilities but also addresses critical concerns of alignment and safety, paving the way for responsible and rapid advancements in artificial intelligence.
Category: Artificial Intelligence
[5] viXra:2409.0068 [pdf] submitted on 2024-09-13 20:56:41
Authors: Maxim Shatskiy
Comments: 4 Pages.
This document describes solution to AutoPET3 Challenge. We show how an ensemble of Unet++ models with EfficientNet-B7 back-bones trained separately on FDG and PSMA data can perform well in this competition. Can a single model beat two specialized models? We see what results of this competition will bring.
Category: Artificial Intelligence
[4] viXra:2409.0063 [pdf] submitted on 2024-09-12 09:25:02
Authors: Jeongik Cho
Comments: 10 Pages.
Classifier gradient penalty GAN is a GAN proposed to perform self-supervised class-conditional data generation and clustering on unlabeled datasets. The classifier gradient penalty GAN's generator takes a continuous latent vector and a categorical latent vector as input and generates a class-conditional data point corresponding to the categorical latent vector. In this paper, we propose to leverage the codebook architecture to improve the performance of classifier gradient penalty GAN. In the proposed architecture, the generator takes the page vector of the codebook corresponding to the index of the categorical latent vector, instead of taking the one-hot categorical latent vector directly. Unlike the codebook used in generative models with vector quantization, the codebook of the proposed architecture is not embedded with the encoder. Instead, the codebook is simply trainable and updated via generator loss like trainable parameters in the generator. The proposed architecture improved the quality of the generated data, class-conditional data generation performance, and clustering performance of the classifier gradient penalty GAN.
Category: Artificial Intelligence
[3] viXra:2409.0056 [pdf] submitted on 2024-09-11 19:55:18
Authors: Maxim Shatskiy
Comments: 4 Pages.
This document describes solution to AutoPET3 Challenge. We show how an ensemble of Unet++ models with EfficientNet-B7 backbones trained separately on FDG and PSMA data can perform well in this competition. Can a single model beat two specialized models? We see what results of this competition will bring.
Category: Artificial Intelligence
[2] viXra:2409.0047 [pdf] submitted on 2024-09-09 17:47:18
Authors: Sing Kuang Tan
Comments: 10 Pages.
In this paper, I am going to propose a new Boolean Structured Variational Autoencoder Deep Learning Network (BSvarautonet) built on top of BSautonet, based on the concept of monotone multi-layer Boolean algebra. Kullback—Leibler (KL) divergence used in traditional Variation Autoencoder has convergence problem and numerical instabilities. Due to the Boolean Structured design of BSautonet, the bottleneck latent space embeddings is naturally distributed in multi-variables Gaussian distribution. By applying a whitening normalization on the latent space, it will transform the latent space to unit Gaussian distribution. Through analysis of the datapoints in latent space and generated MNIST digit images, it has shown that it has all the properties of variational autoencoder. The BS autoencoder is a masked noise denoising model, therefore it can acts like a diffusion model to incrementally generate a digit image from a noisy one through repeated applications of the autoencoder model.
Category: Artificial Intelligence
[1] viXra:2409.0018 [pdf] submitted on 2024-09-04 20:17:28
Authors: R. Peeyoos
Comments: 26 Pages. (Note by viXra Admin: Author's first name is required)
With advancements in large language models (LLMs) and multimodal AIs capable of code, media, automation, the realization of artificial general intelligence (AGI) is increasingly plausible. As the potential for achieving sentient AGI within the coming decades grows, implementing effective safety measures to align AGI with human interests becomes crucial. Current AGI safety strategies primarily focus on hardware, coding, and mathematical constraints, but these may not be sustainable in the long term. As AGI evolves, it could bypass or overcome these limitations. This paper introduces a novel approach to AGI alignment by avoiding traditional safety measures in areas where AGI is inherently strong. Instead, it proposes establishing a symbiotic relationship between humans and AGI, leveraging human strengths and AGI's vulnerabilities. This approach aims to ensure AGI's benevolence by choice, reducing its motivation to act against humanity and providing a more reliable long-term solution compared to conventional strategies that enforce compliance.
Category: Artificial Intelligence
| Third-Party Links: |
|