Devising model compression methodologies for sustainable convolutional neural networks

Abstract

Convolutional Neural Networks (CNNs) have significantly advanced various fields in computer vision such as image classification, object detection and image segmentation. However, deploying them in real-world applications presents significant challenges. Implementing CNNs on devices with limited computational resources, like mobile phones, edge devices and embedded systems, can be difficult due to memory and processing constraints. This research focuses on developing CNN model compression strategies, with the goal of reducing the number of model parameters, without significant loss in performance. When such compact CNN models are deployed on devices with limited resources, it not only enhances energy efficiency but also ensures faster inference times especially for time-sensitive applications. newlineThe first research work, quotDCompressNet: An efficient approach to escalate the speed of training Convolution Neural Networks through Data Compression and Network Pruningquot addresses the challenge of speeding up training as well inference times of complex CNN architectures. The proposed system DCompressNet identifies and eliminates non-critical data points during training using a granularity-based aggregation approach. Moreover, to address the challenge of deploying complex CNN models on resource-constrained devices, a model compression technique called soft filter pruning is employed, that overcomes limitations of traditional hard filter pruning techniques. Experimental validation on two benchmark datasets CIFAR-10 and CIFAR-100, demonstrates a significant reduction in Floating Point Operations per Second (FLOPS) of four CNN models without significant loss of accuracy, making it a promising solution for acceleration. newline newline