Fischer/etal/2022a: A Unified Framework for Assessing Energy Efficiency of Machine Learning

Bibtype Inproceedings
Bibkey Fischer/etal/2022a
Author Fischer, Raphael and Jakobs, Matthias and Mücke, Sascha and Morik, Katharina
Ls8autor Fischer, Raphael
Jakobs, Matthias
Morik, Katharina
Mücke, Sascha
Title A Unified Framework for Assessing Energy Efficiency of Machine Learning
Booktitle Proceedings of the ECML Workshop on Data Science for Social Good
Abstract State-of-the-art machine learning (ML) systems show exceptional qualitative performance, but can also have a negative impact on society. With regard to global climate change, the question of resource consumption and sustainability becomes more and more urgent. The enormous energy footprint of single ML applications and experiments was recently investigated. However, environment-aware users require a unified framework to assess, compare, and report the efficiency and performance trade-off of different methods and models. In this work we propose novel efficiency aggregation, indexing, and rating procedures for ML applications. To this end, we devise a set of metrics that allow for a holistic view, taking both task type, abstract model, software, and hardware into account. As a result, ML systems become comparable even across different execution environments. Inspired by the EU’s energy label system, we also introduce a concept for visually communicating efficiency information to the public in a comprehensible way. We apply our methods to over 20 SOTA models on a range of hardware architectures, giving an overview of the modern ML efficiency landscape.
Year 2022

  • Privacy Policy
  • Imprint