Optimizing Machine Learning Models for Resource-Constrained Environments: A Comprehensive Study of Compression and Quantization Across CIFAR-10, MNIST, and Fashion MNIST
Prachi Garg, Avinash Sharma, Karan Karan, Rini Saxena, Parul Goyal, Tarkeshwar Barua · 2025
Integration of machine learning (ML) in resource-constrained environments, such as Internet of Things (IoT) devices, embedded systems, and mobile applications, has become one of the crucial focus areas of modern research and development. Each of these environments has unique challenges—be it insufficient computational power, memory, or energy resources— which necessitates developing optimized ML models that can execute high performance under these constraints. This paper addresses advanced compression and quantization techniques applied to ML models trained on widely used datasets such as CIFAR-10, MNIST, and Fashion MNIST. Through an extensive evaluation process, we analyse key performance metrics, including accuracy, latency, model size, and energy efficiency to find a trade-off among the adopted techniques.The results show large reductions in latency, up to 70% better, and model size, with more than 90% storage saving, can be obtained with only small losses in accuracy,making these techniques applicable in practice in edge devices. More importantly, the study highlights the potential of hybrid optimization strategies by combining compression and quantization to further improve the performance and adaptability of the models with techniques such as adaptive quantization that dynamically adjusts to real-world complexity. This work also identifies gaps in current research and suggests dataset-specific pipelines for actionable insights, contributing toward efficient ML system development for a wide range of resource-constrained domains. Overall, these findings put into perspective the transformative potential these techniques hold for enabling sophisticated ML model deployment where traditional models become infeasible due to resource constraints.