A comparative study of neural network pruning strategies for industrial applications

Abstract

Introduction: In recent years, Deep Learning (DL) and Artificial Neural Networks (ANNs) have transformed industrial applications by providing automation in complex tasks such as anomaly detection and predictive maintenance. However, traditional DL models often need significant computational resources, making them unsuitable for resource-constrained edge devices. This paper explores the potential of sparse ANNs to address these challenges, focusing on their application in industrial settings. Methods: We perform an experimental comparison of pruning techniques, including the Pre-Training, In-Training, Post-Training, and SET method, applied to the VGG16 and ResNet18 architectures, and conduct a systematic analysis of pruning methodologies alongside the effects of varying sparsity levels, to analyze their efficiency in anomaly detection and object classification tasks. Key metrics such as training accuracy, inference time, and energy consumption are analyzed to assess the feasibility of deploying sparse models on edge devices. Results and discussion: Our results demonstrate that sparse ANNs, particularly when pruned using the SET method, achieve energy savings without compromising accuracy, making them suitable for industrial applications. This work highlights the potential of sparse neural networks to boost sustainability and efficiency in industrial environments, paving the way for their large adoption in edge computing scenarios.