North Carolina State University predicts AI energy costs

Researchers at North Carolina State University have developed a novel method to predict the computational and energy costs associated with updating deep learning artificial intelligence models. According to Jung-Eun Kim, assistant professor of computer science, this new technique called RESQUE allows users to compare the initial dataset used to train a model with the new dataset used for updates, estimating the costs required for retraining.

This innovation has significant implications for AI sustainability as it enables informed decisions about when to update models and how to budget computational resources. Kim collaborated with graduate student Vishwesh Sangarya on the project, which will be presented at the Thirty-Ninth Association for the Advancement of Artificial Intelligence Conference. The research aims to make deep learning models more sustainable by reducing the need for frequent retraining from scratch, a process that demands substantial computational power and energy.

Introduction to Sustainable AI Models

The development and implementation of deep learning models have become increasingly prevalent in various industries, with a growing focus on sustainability. As these models are updated to accommodate new tasks or changes in data, the computational resources and energy consumption required can be substantial. Researchers have developed a novel method, called Representation Shift QUantifying Estimator (RESQUE), to predict the computational and energy costs associated with updating deep learning models. This approach enables users to make informed decisions about when to update their models, ultimately contributing to more sustainable AI practices.

The process of updating deep learning models involves retraining existing models or training new ones from scratch. However, retraining an existing model is often more cost-effective than training a new one, as it requires less computational power and energy. RESQUE allows users to compare the initial dataset used to train a model with the new dataset that will be used for updates, estimating the computational and energy costs associated with the update process. The estimated costs are presented as a single index value, which can be compared with various metrics such as epochs, parameter change, gradient norm, carbon emissions, and energy consumption.

The development of RESQUE is crucial in the context of sustainable AI, as it provides a deeper understanding of the costs associated with deep learning models across their entire life cycle. By predicting the computational and energy costs required for updates, users can budget their resources more effectively, predict the time required for updates, and make informed decisions about model reusability. The long-term implications of this research are significant, as it has the potential to contribute to the development of more sustainable AI models that are not only dynamic but also environmentally friendly.