Machine learning, in its various tasks from fitting to inference, can be highly energy intensive and raises growing environmental concerns. This situation inspired different initiatives fostering a more frugal, greener AI.
Beyond the implementation of good practices, it appears pivotal for researchers and data engineers to gather an empiric knowledge of energy consumption per task, data profile, hardware environment, to improve calibration strategies, promote less thirsty algorithms, and save money. To do so, the Research Domain ‘Responsible AI’ proposes an agnostic and modular format, both object-oriented and tabular, which could help as a common ground in various software libraries, devices and digital ecosystems, to achieve the open data promise in the power consumption of machine learning.
The format comes with an extensive definition in JSON schema, integrity tests and tabular conversion tooling. The public GitHub repository contains Python source code under CC 4.0 license to know more about the proposal and participate to roadmap discussions.