Title: Supervised Deep Learning for Real-Time Big Data Streams in IoT Networks: A Review of Latency-Accuracy Trade-offs
Authors: Virendra Tank
Volume: 10
Issue: 7
Pages: 202-209
Publication Date: 2026/07/28
Abstract:
The rapid expansion of Internet of Things (IoT) networks has produced continuous, high-velocity data streams that demand supervised deep learning models capable of making accurate predictions within strict real-time latency budgets. Unlike static big-data analytics, real-time IoT streaming applications - such as autonomous navigation, video surveillance, and industrial anomaly detection - must jointly optimize two often conflicting objectives: minimizing end-to-end latency and maximizing predictive accuracy. This review paper systematically examines the latency-accuracy trade-off in supervised deep learning for real-time big data streams in IoT networks. It surveys four major families of optimization strategies: model compression (pruning, quantization, and knowledge distillation), dynamic and early-exit inference architectures, edge-cloud cooperative offloading, and streaming-aware evaluation and forecasting methods. Representative studies are analyzed and compared in terms of latency reduction, accuracy impact, and suitability for different IoT application domains. The paper further discusses open challenges, including the mismatch between offline accuracy metrics and real-world streaming performance, resource constraints on edge hardware, and network variability. It concludes with future research directions such as streaming-native model architectures, reinforcement-learning-based adaptive offloading, and standardized streaming-accuracy benchmarks for IoT deep learning systems.