Convergence of Edge Computing and Machine Learning for Real-Time Intelligent Decision-Making Systems: A Structured Review of Architectures, Optimisation Techniques, and Deployment Challenges

Arindam Sen, Sneha Bhattacharya

Abstract


ABSTRACT
Real-time decision-making systems such as autonomous vehicles, industrial control loops, and remote health monitors must convert streaming sensor data into actions within deadlines that range from a few milliseconds to a few hundred milliseconds. Cloud-centric machine learning (ML) pipelines struggle to meet such deadlines because of wide-area network delay, bandwidth cost, and exposure of raw data. Edge computing relocates computation close to the data source, but running modern ML models on resource-constrained edge nodes raises new problems of model size, energy consumption, and distributed training. This review-based study synthesises peer-reviewed literature published between 2015 and 2021 on the convergence of edge computing and ML, using a four-layer analytical framework that covers reference architecture, model optimisation, collaborative inference, and federated learning. The reported evidence shows that pruning and quantisation pipelines can shrink deep networks by 35 to 49 times, that partitioned inference can reduce average latency by about 3.1 times, and that federated averaging can cut communication rounds by 10 to 100 times. No single technique is sufficient: the dominant latency term shifts from the network to model inference as computation moves to the edge, so architecture, compression, and learning strategy must be co-designed. The paper concludes with limitations, practical applications, and priorities for future research, including standardised benchmarks, security, and on-device continual learning.

KEYWORDS: Edge computing, Machine learning, Edge intelligence, Real-time decision-making, Federated learning, Model compression, TinyML, Computation offloading, Internet of Things, Collaborative inference

Full Text:

PDF 103-117