A standard reporting system for the environmental impact of machine learning

Abstract

The growing demand of compute and resources required for developing machine learning models has led to an increased adverse impact on the environment. However, there is a lack of data concerning the environmental footprint of machine learning models available in the public domain. Even when data is available, important parameters such as water consumption are ignored. This paper aims to provide a standardized benchmark to report the environmental impact of individual machine learning models in terms of energy use, water consumption and carbon footprint. The proposed documentation system, referred to as the EnvCard, is intended to be an analogue to the model card for model reporting, helping stakeholders make more resource aware decisions. EnvCards are intended to be a stepping stone towards increasing transparency about the unintended consequences of the accelerated development of Artificial Intelligence technologies.