Generating Energy-Efficient Neural Networks

Research Overview

In generating energy-efficient neural networks, specfically convolutional neural networks, I will take a given neural network architecture as input, and return a new nerual network architecutre optmized for energy-efficiency. The internal process by which the returned network is generated is through an iterative neural architecture searching process, optimizing for energy usage and runtime instead of accuracy and error rate.

Gap

While there are papers discussing the development of AI toolting that will model the energy requirements for a particular stage of the model, think the Paleo program by Hang Qi. et al and the NeuralPower program by Ermao Cai. et al, these programs only work on a single model. To this end, I am interested in taking these mathematical models derived for Paleo and NeuralPower, and using them to iteratively generate, at least potentially, more energy efficient AI. Interestingly, there already exists a method by which to generate optimized AI models, these are called Neural Architecture Searching algorithms. But, these searching algorithms are mostly used for taking a dataset, such as image recognition, and generating a model with a minimal error rate for that particular dataset. For the specific algorithm I will be using, I can swap out this performance predictor, used in the iterative optimizing process, for the Paleo and NeuralPower models--this is the gap I am attacking. To my knowledge, there is no program that will iteratively generate a more energy-efficient model based upon a given architecture.