AI Innovation: New Tool Forecasts Costs for Sustainable Model Updates

Lifestyle: January 20, 2025: Artificial intelligence (AI) continues to shape the modern world, but its environmental and computational costs remain a concern. Recognizing this challenge,

 

researchers at North Carolina State University have developed a groundbreaking tool to predict the costs of updating AI models. This new method could play a pivotal role in improving the sustainability of deep learning technologies.

 

 

Deep learning models, a subset of AI, often require updates to stay relevant. These updates, known as model retraining, are essential when tasks evolve or when new data becomes available. While updates ensure accuracy and functionality, they come with significant computational and energy demands.

 

“There have been studies that focused on making deep learning model training more efficient,” explains Jung-Eun Kim, assistant professor of computer science at NC State and corresponding author of the study. “However, over a model’s life cycle, it will likely need to be updated many times. Our work shows that retraining an existing model is much more cost-effective than training a new one from scratch.”

Kim highlights that understanding the costs of updates—both computational and environmental—is crucial to planning and achieving sustainability in AI systems.

 

 

To meet this challenge, the research team developed the REpresentation Shift QUantifying Estimator (RESQUE), a technique that predicts the computational and energy demands of updating AI models. RESQUE compares an AI model's original dataset to the new dataset required for updates. Based on this comparison, it provides a detailed estimate of the costs associated with retraining.

 

These costs are summarized in a single index value, which users can interpret through metrics such as epochs (iterations in training), parameter changes, and gradient norms (measures of training effort). Additionally, RESQUE translates these metrics into practical energy and carbon impact values, estimating:

 

 

Carbon emissions in kilograms

This dual approach provides a clearer understanding of the environmental consequences of AI updates, helping users make informed decisions.

 

 

Two primary changes necessitate AI model updates:

Task Shifts: When an AI model’s role evolves, such as expanding from recognizing traffic symbols to identifying vehicles and pedestrians.
Distribution Shifts: When the data the model uses changes, such as transitioning to new data formats or types.

 

Both shifts require computational effort, but with RESQUE, AI practitioners can anticipate the effort needed and plan updates efficiently.

“We found that the RESQUE predictions aligned very closely with the real-world costs of conducting deep learning model updates,” says Kim. “All our experimental findings also confirm that retraining a model demands far fewer resources than building a new one.”

 

 

Beyond budgeting resources for updates, RESQUE sheds light on the long-term sustainability of AI models. By quantifying the energy and carbon footprints of model updates, the methodology supports efforts to make AI greener and more sustainable.

 

“In the bigger picture, this work offers a deeper understanding of the costs associated with deep learning models across their entire life cycle,” notes Kim. “If we want AI to be viable and useful, these models must be not only dynamic but sustainable.”

 

 

The findings will be presented at the Thirty-Ninth Association for the Advancement of Artificial Intelligence (AAAI) Conference on Artificial Intelligence in Philadelphia from February 25 to March 4. The study, titled “RESQUE: Quantifying Estimator to Task and Distribution Shift for Sustainable Model Reusability,” is authored by Vishwesh Sangarya, a graduate student at NC State, alongside Kim and the research team.

As AI becomes more integrated into everyday life, tools like RESQUE are essential to ensuring that advancements in technology align with environmental and sustainability goals. This innovative approach not only benefits researchers and developers but also contributes to a greener future for AI.

 

Source: NC State University: New Method Forecasts Computation, Energy Costs for Sustainable AI Models Jan 13, 2025

 

WNCTimes

 

Image: WNCTimes