Using Visualisation to Unpack, Reveal, Challenge and Transform the Ecological Impact of ML Processes

The training and fine-tuning of Machine Learning (ML) models is costly to the environment in terms of computational power (energy) and ultimately carbon emissions (Bender et al 2021, Schwartz et al 2020, Strubell et al 2019). For example, roughly 15% of Google’s energy consumption in 2021 was dedicated to AI tasks (Patterson et al 2022). With the ongoing increase in Machine Learning (ML) in both model numbers as well as their achieved complexity, its environmental impact is adding up to be significant. While the ML and HCI communities has been identifying such critical environmental issues, there is still a lack of comprehensive methodologies to support system designers and developers in taking a sustainability-by-design approach where they understand, measure, and mitigate such issues. For instance, the few existing tools that calculate carbon emission costs (Anthony et al 2020, Lacoste et al 2019, Lannelongue 2021) largely focus on technical measurements, largely overlooking if and how they are used by practitioners in the field.

The examination of eco-feedback technology’s impact on behaviour change has a long history in Human-Computer Interaction (HCI) and Visualization research. Such research has often focused on domestic settings for instance understanding how people react to smart meter displays or custom energy visualisations (e.g., Costanza et al 2012). Parallel to the state-of-the-art in ML carbon tracking, early such work also was centred around the tools themselves, only later maturing to appreciate the more complex socio-technical issues that define if and how eco-feedback impacts behaviour. For instance, early studies documented a correlation between eco-feedback and feelings of guilt (Prost et al. 2015), yet more recent work has since focused on empowering messages such as the encouragement of collective action (Panagiotidou et al 2023; 2024, Bird and Rogers 2010).

This PhD will draw from research on sustainable HCI, behaviour-change and critical data visualisation to examine how to best communicate ML’s potential environmental effects. Specifically the student will investigate current sustainability-oriented practices within ML teams and strive towards developing a design toolkit for supporting more sustainable-oriented decision-making. The project may include activities such as conducting interviews and observations, running co-design workshops as well as developing visualisations of carbon emissions of machine learning models to act as a tools meant for eco-feedback.

Anthony, L. F. W.; Kanding, B.; Selvan, R. 2020. Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models. arXiv.

Bender E. M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21). Association for Computing Machinery, New York, NY, USA, 610–623.

Bird, J. and Rogers, Y., 2010, May. The pulse of tidy street: Measuring and publicly displaying domestic electricity consumption. In workshop on energy awareness and conservation through pervasive applications (Pervasive 2010).

Costanza E., Sarvapali D. Ramchurn, and Nicholas R. Jennings. 2012. Understanding domestic energy consumption through interactive visualisation: a field study. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing (UbiComp ’12). Association for Computing Machinery, New York, NY, USA, 216–225.

Lacoste, A., Luccioni, A., Schmidt, V. and Dandres, T., 2019. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv: 1910.09700.

Lannelongue, L., J. Grealey, M. Inouye, 2021. Green Algorithms: Quantifying the Carbon Footprint of Computation. Advanced Science 8, 2100707.

Open AI. 2018. AI and compute.

Panagiotidou G., Enrico Costanza, Michael J. Fell, Farhan Samanani, and Hannah Knox. 2023a. Supporting Solar Energy Coordination among Communities. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 7, 2, Article 71 (June 2023), 23 pages.

Panagiotidou G., Enrico Costanza, Michael J. Fell, Kyrill Potapov, Sonia Mugambi, Farhan Samanani, and Hannah Knox. 2024. SolarClub: Supporting Renewable Energy Communities through an Interactive Coordination System. CHI’24

Prost S., E. Mattheiss, and M. Tscheligi. 2015. ‘From Awareness to Empowerment: Using Design Fiction to Explore Paths towards a Sustainable Energy Future’, in Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing, Vancouver BC Canada: ACM, Feb. 2015, pp. 1649–1658.

Schwartz R., Jesse Dodge, Noah A. Smith, and Oren Etzioni. 2020. Green AI. Communications. ACM 63, 12 (December 2020), 54–63.

Strubell E., Ananya Ganesh, and Andrew McCallum. 2019. Energy and Policy Considerations for Deep Learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645–3650, Florence, Italy. Association for Computational Linguistics

Project ID