Deep learning audio development costs far more energy than training
Exposing the Cost of Deep Learning Audio Development
Machine LearningArtificial IntelligenceSound
Summary
Usually, people look at how much energy deep learning uses when training or using a model, but they often forget about the time spent developing the model. The authors studied how much energy is used during the trial-and-error stage of building audio deep learning projects. They found this development phase can use 3 to 256 times more energy than just training the final model. This suggests we should measure and report energy use across the whole project, not just training.
What this means in practice
- •For machine learning engineers: Estimate energy costs more accurately by including prototype and experiment phases in deep learning audio projects.
- •For data center operators: Plan power usage and carbon footprint more effectively by accounting for energy-intensive development activities beyond model training.
Authors
Constance Douwes, Paul Magron, Romain Serizel
Abstract
The environmental impact of deep learning has attracted increasing attention over the past decade. Existing studies mainly focus on the energy and carbon emissions of model training and inference, while the whole development phase is often overlooked. Yet, architecture prototyping and intensive experiments are conducted during this stage, which is highly energy-demanding. In this article, we propose a methodology to estimate these costs, based on activity logs from the Grid5000 shared computing platform used by the LORIA laboratory. As a case-study, we focus on audio projects developed in the Multispeech research team. We evaluate the overall energy cost of four projects, and we compare them to those of training the reported models. Our results show that the energy required for the development phase is 3 to 256 times greater than that required to train the best-performing model alone. These results advocate for a more systematic reporting of energy consumption across the entire life cycle of deep learning-based audio projects.