Abstract
Abstract Deep Convolutional Neural Networks (CNNs) have gained immense popularity over the past decade due to their exceptional performance, especially in imaging applications. However, the primary focus has been on improving model accuracy, often overlooking the significant environmental and computational costs associated with it. Pruning, a model compression technique, has been popularly used to reduce model size and computational complexity. Despite rapid growth of interest in this topic, research which comprehensively studies the energy consumption and targets energy efficiency using structured pruning is to date still missing. In this paper, we propose a simple, novel and energy-efficient structured pruning methodology resulting in energy, memory footprint and parameter reduction. Using the above methodology, an energy reduction of ranging from 6.49% to 58.58% is achieved, depending on the architecture and dataset. For VGG16 on CIFAR-10, the proposed methodology brings about 58% inference energy reduction, 89% reduction in memory footprint and 54% latency reduction for a negligible drop ( $$\sim 1\%$$ ) in accuracy. The one-shot version of the energy-aware pruning algorithm achieves an impressive 74.64% reduction in training energy for VGG16 on CIFAR-10, leading to substantial additional energy savings. We also provide practical insights into the energy-accuracy tradeoff and encourage researchers to choose a suitable approach for pruning based on the required accuracy rather than best possible accuracy for a particular application to ensure sustainable and Green AI.