Analysis of Energy-Efficient LCRM Optimization Algorithm in Computer Vision-based CNNs

Enhancing the energy efficiency of deep learning models, such as Convolutional Neural Networks (CNNs), is important in enabling sustainable AI. This research presents a comprehensive analysis of the Energy-Efficient Layer-wise complexity reduction (LCRM) optimization algorithm in the context of computer vision-based CNNs. The study focuses on evaluating the algorithm's effectiveness in reducing CNN's energy consumption without compromising the overall performance of the neural network. The work is evaluated across optimizing standard CNN models, including AlexNet, VGG-9, and U-Net, spanning applications like image classification, optical character recognition (OCR), and image segmentation. The methodology also analyzes the comparison of LCRM with state-of-the-art (SOTA) optimization techniques like quantization pruning and clustering, emphasizing its role in minimizing energy consumption. The outcome of our research is beneficial for industries to quantify the level of sustainability provided by different energy-efficient solutions for optimizing AI models. The implementation of the energy CNN model analysis is available at: https://github.com/hana-an/LCRM/blob/main/docs/examples/model-energy.ipynb