DYNAMIC MULTI-OBJECTIVE PRUNING OF DEEP NEURAL NETWORKS FOR ENERGY-EFFICIENT IOT EDGE COMPUTING

Abstract

Deep Neural Networks (DNNs) are severely restricted from being deployed on Internet of Things (IoT) edge devices due to hardware constraints such as limited memory, low power and limited battery life. Model compression techniques reduce the storage overhead of deep learning models by pruning. However, conventional approaches do not take into account the changing energy levels within real-world IoT environments. To address this issue, this work presents a multi-objective evolutionary approach using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) for dynamically pruned convolutional neural networks (CNNs). Structural filters and dense neurons are represented using binary encoding to allow for both maximal classification accuracy and minimal number of active parameters. Experiments were carried out using the UCI Human Activity Recognition (HAR) dataset, resulting in a clear Pareto Front of non-dominated solutions. The resulting pruned networks produced near 50% fewer active parameters while maintaining a high accuracy of 86.4%. Extreme levels of model compression achieved up to 66% of the baseline accuracy with very little energy consumption. This provides a powerful free solution for Edge AI to dynamically balance computation on resource-constrained IoT systems.