Compression and acceleration of convolution neural network: a Genetic Algorithm based approach

Abstract

Genetic Algorithm (GA) is a meta-heuristics search and optimization approach which we have utilized for fine-tuning the standard deep learning models for reducing the storage space with improvement in inference time. The pre-trained models have been widely acclaimed as the best choice of CNN for various image classification problems in different domains, but they require huge storage space and it causes a problem for deploying these models on mobile or edge devices. As these devices are constrained with limited memory and computational power. In this paper, a novel GA based method has been proposed which compresses and accelerates the CNN models so that these models can be easily deployed on edge devices. Extensive computer simulations have been conducted on widely used models such as AlexNet, VGG16, SqueezeNet, and ResNet50. We used benchmark datasets such as MNIST, CIFAR-10, and CIFAR-100 to determine the performance. Results reveal that storage space of the AlexNet model was reduced by 87.5%, 86.55% and 86.16% on MNIST, CIFAR-10 and CIFAR- 100 datasets respectively, while VGG16, ResNet50, and SqueezeNet have been compressed by nearly 91%, 78%, 38% respectively. From the results, it has been noticed that there is a significant improvement in inference time of around 35% in AlexNet, 9% in SqueezeNet, 73% in ResNet50, and 80% in VGG16. This improvement is noticed mainly because of the fine-tuning of the deep learning models using the GA. Overall, the proposed GA-based method showed outstanding performance and it motivates research and practitioners to explore it further. © 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.