NeurstrucEnergy: A bi-directional GNN model for energy prediction of neural networks in IoT

Abstract

A significant demand rises for energy-efficient deep neural networks to support power-limited embedding devices with successful deep learning applications in IoT and edge computing fields. An accurate energy prediction approach is critical to provide measurement and lead optimization direction. However, the current energy prediction approaches lack accuracy and generalization ability due to the lack of research on the neural network structure and the excessive reliance on customized training dataset. This paper presents a novel energy prediction model, NeurstrucEnergy. NeurstrucEnergy treats neural networks as directed graphs and applies a bi-directional graph neural network training on a randomly generated dataset to extract structural features for energy prediction. NeurstrucEnergy has advantages over linear approaches because the bi-directional graph neural network collects structural features from each layer's parents and children. Experimental results show that NeurstrucEnergy establishes state-of-the-art results with mean absolute percentage error of 2.60%. We also evaluate NeurstrucEnergy in a randomly generated dataset, achieving the mean absolute percentage error of 4.83% over 10 typical convolutional neural networks in recent years and 7 efficient convolutional neural networks created by neural architecture search. Our code is available at https://github.com/NEUSoftGreenAI/NeurstrucEnergy.git.

Keywords

• Internet of things
• Neural network energy prediction
• Graph neural networks
• Graph structure embedding
• Multi-head attention