Paper Key : IRJ************118
Author: Abdulaziz J. Almarashi
Date Published: 09 Mar 2024
Abstract
This paper presents a study on the optimization of Convolutional Neural Networks (CNNs) using Taylor series pruning, a technique that identifies and eliminates less significant convolution filters in the network. The importance of CNN pruning lies in its ability to reduce the computational cost and memory usage, thereby increasing the efficiency and speed of the network, which is crucial for real-time applications. The experiments were conducted on the CIFAR-10 dataset, a standard benchmark in the field of image classification. Prior to pruning, the network achieved an accuracy of 90.5%. Post-pruning, the accuracy slightly decreased to 88.1%. However, this minor loss in accuracy was counterbalanced by a substantial gain in computational efficiency. The pruning process resulted in a 25% reduction in the number of convolution filters, leading to a significant improvement in inference speed. The pruned network was deployed on the NVIDIA Jetson Nano device using MATLAB GPU Coder. The results demonstrate the effectiveness of Taylor series pruning in optimizing CNNs for edge computing devices without significantly compromising classification performance. This study contributes to the ongoing efforts to make deep learning models more efficient and accessible for real-world applications.
DOI LINK : 10.56726/IRJMETS50054 https://www.doi.org/10.56726/IRJMETS50054