Scalable Deep Learning Frameworks for Distributed Computing Environments

Abstract

The rapid growth of large-scale visual data generated from intelligent surveillance systems, Internet of Things (IoT) enabled cameras, and smart monitoring environments has created significant challenges for efficient Deep Learning (DL) model training in distributed computing infrastructures. These environments require scalable frameworks capable of handling continuous high-volume image streams while ensuring fast training, optimal resource utilization, and minimal communication overhead across Multi-GPU Systems. To address these challenges, this research proposes a scalable distributed DL framework designed specifically for large-scale image-based analytics in IoT and intelligent monitoring applications. The proposed framework integrates data-parallel training with an optimized communication mechanism based on the Seven-Spot Ladybird Optimized Intelligent VariationalAutoencoder (SSLO-IntVAE), enabling efficient synchronization across distributed GPUs while reducing inter-node communication costs. In addition, an optimized data pipeline is incorporated to minimize input/output bottlenecks and improve throughput during large-scale training processes. A publicly available large-scale image dataset is utilized as a representative benchmark of visual data generated in IoT-based surveillance and monitoring systems. The framework is implemented and tested on a multi-GPU cluster using DL models trained in a distributed environment. Experimental results demonstrate significant improvements in 93.18% of precision, 92.53% of recall, and 96.10% of F1-score. The proposed framework provides a robust solution for large-scale visual data processing, supporting intelligent decision-making in distributed IoT and surveillance systems.