A binary particle swarm optimization-based pruning approach for environmentally sustainable and robust CNNs

Abstract

Deep Convolutional Neural Networks (CNNs), continue to demonstrate remarkable performance across various tasks. However, their computational demands and energy consumption present significant drawbacks, restricting their practical deployment and contributing to a substantial carbon footprint. This paper addresses this challenge by proposing a novel method named Binary Particle Swarm Optimization Layer Pruner (BPSO-LPruner), aimed at achieving substantial computational reduction and mitigating environmental impact during CNN inference. BPSO-LPruner utilizes a constrained Binary Particle Swarm Optimization for CNN layer pruning, integrating a masked-bit strategy and a new population initialization strategy to enhance search performance. We illustrate the effectiveness of our method in reducing model computational costs and carbon footprint emissions while improving performance across multiple models (VGG16, VGG19, DenseNet-40, ResNet18, ResNet20, ResNet34, ResNet44, ResNet56, ResNet110, ResNet50, and MobileNetv2) and diverse datasets (CIFAR-10, CIFAR-100, Tiny-ImageNet, COVID-19 X-ray dataset). Promising results underscore the performance of the proposed method. Additionally, we demonstrate that layer pruning yields benefits beyond enhanced computational performance. Our experimentation reveals that BPSO-LPruner enhances the model's reliability and robustness by effectively addressing variations in input data, inherent ambiguity in model parameters, and adversarial images.