The computations required for deep learning research have been doubling every few months, resulting in an estimated 300,000x increase from 2012 to 2018. This trend has led to unprecedent success in a range of AI tasks. In this talk I will discuss a few troubling side-effects of this trend, touching on issues of lack of inclusiveness within the research community, challenges in adoption of the technology, and an increasingly large environmental footprint.
I will then present Green AI – an alternative approach to help mitigate these concerns. Green AI is composed of two main ideas: increased reporting of computational budgets, and making efficiency an evaluation criterion for research alongside accuracy and related measures. I will discuss these two ideas, presenting recent efforts and open questions.

This is joint work with Dallas Card, Jesse Dodge, Oren Etzioni, Suchin Gururangan and Noah A. Smith.

The carbon footprint of (scientific) computing is a growing source of concern, and there is widespread interest in the research community; so why is this issue still often overlooked? One reason is the perceived barrier to entry to green computing, both in terms of time and resources needed to estimate and reduce the environmental impact of our work. Tackling this issue and reducing frictions for scientists is what motivated the Green Algorithms project. We will discuss what we learned along the way, what still needs to be done as well as simple and impactful ways by which scientists and institutions can make their research more sustainable.

In this talk we examine peer-reviewed studies which estimate ICT's current share of global greenhouse gas (GHG) emissions to be 1.8-2.8% of global GHG emissions. Our findings indicate that published estimates all underestimate the carbon footprint of ICT, possibly by as much as 25%, by failing to account for all of ICT's supply chains and full lifecycle (i.e. emissions scopes 1, 2 and fully inclusive 3). Adjusting for truncation of supply chain pathways, we estimate that ICT's share of emissions could actually be as high as 2.1-3.9%. We explore the argument for and against the role of efficiency gains and green energy in offsetting ICTs global carbon footprint, and the potential for efficiency gains to lead to rebound effects (c.f. Jevons Paradox). Whatever assumptions analysts take, they agree that ICT will not reduce its emissions without a major concerted effort involving broad political and industrial action. We make specific recommendations and pose a set of challenges for those using heavy computation in their research practice.

The talk is based on our public report: https://arxiv.org/abs/2102.02622