Architecture optimization techniques for Convolutional Neural Networks: further experiments and insights

Abstract

In this paper, we have researched implementing convolutional neural network (CNN) models for devices with limited resources, such as smartphones and embedded computers. To optimize the number of parameters of these models, we studied various popular methods that would allow them to operate more efficiently. Specifically, our research focused on the ResNet-101 and VGG-19 architectures, which we modified using techniques specific to model optimization. We aimed to determine which approach would work best for particular requirements for a maximum accepted accuracy drop. Our contribution lies in the comprehensive ablation study, which presents the impact of different approaches on the final results, specifically in terms of reducing model parameters, FLOPS, and the potential decline in accuracy. We explored the feasibility of implementing architecture compression methods that can influence the model’s structure. Additionally, we delved into post-training methods, such as pruning and quantization, at various model sparsity levels. This study builds upon our prior research [1] to provide a more comprehensive understanding of the subject matter at hand.