FedPromo: Federated Lightweight Proxy Models at the Edge Bring New Domains to Foundation Models

A new approach utilizing multi-head attention and the Inception v3 model has been developed for the automatic detection of cardiomegaly through X-ray images. This method integrates deep learning tools and attention mechanisms, enhancing the accuracy and efficiency of diagnosing cardiovascular diseases by leveraging a robust data collection phase and preprocessing techniques to improve image quality.

A recent study highlights the importance of automated crack detection in preserving cultural heritage artifacts through the use of semantic segmentation techniques. The research focuses on evaluating various U-Net architectures for pixel-level crack identification on statues and monuments, utilizing the OmniCrack30k dataset for quantitative assessments and real-world evaluations.

A new framework called Coupled Physics-Gated Adaptation (C-PGA) has been introduced to predict photochemical conversion in complex 3D-printed objects, utilizing a large dataset of optically printed specimens. This innovative approach addresses the limitations of conventional vision models in understanding the coupled interactions of optical and material physics that influence chemical states.

HVAdam, a novel full-dimension adaptive optimizer, has been introduced to address the performance gap between adaptive optimizers like Adam and non-adaptive methods such as SGD, particularly in training large-scale models. The new optimizer features continuously tunable adaptivity and a mechanism called incremental delay update (IDU) to enhance convergence across diverse optimization landscapes.

A new approach to text-image-to-video (TI2V) generation has been introduced, known as FlexTI2V, which allows for flexible visual conditioning without the need for extensive training. This method enhances the capabilities of text-to-video (T2V) models by incorporating arbitrary images at various positions, utilizing a novel random patch swapping strategy during the denoising process.

AraFinNews has been introduced as the largest publicly available Arabic financial news dataset, featuring 212,500 article-headline pairs from 2015 to 2025, aimed at enhancing Arabic financial text summarization using large language models (LLMs). The dataset serves as a benchmark for evaluating language understanding and generation in financial contexts, particularly through transformer-based models like mT5, AraT5, and FinAraT5.

ParaBlock is a novel approach to federated learning that enhances communication efficiency by establishing parallel threads for communication and computation, addressing the challenges faced by resource-constrained clients when training large language models (LLMs). This method theoretically matches the convergence rate of standard federated block coordinate descent methods.

Recent research has analyzed the use of momentum in decentralized and federated optimization, particularly in the context of Federated Learning (FL). The study reveals that while momentum can help mitigate statistical heterogeneity, it does not guarantee convergence under unbounded conditions, especially with cyclic client participation. The findings indicate that decreasing step-sizes do not improve convergence outcomes, leading to a constant value influenced by initialization and heterogeneity bounds.