Tiny machine-learning models, or none at all: a provocation on the environment impact of machine learning
Laura U. Marks
■ Abstract
Machine learning (ML) is often touted to improve the efficiency of networked digital media, but that small gain is overwhelmed by the enormous carbon, water, and land footprints of data centers and ML-ready devices. Drawing on my past work in translating engineering research for broader publics, I survey the high electricity consumption of ML applications in training and inference, focusing on electricity-intensive image generation. The solutions I share are developed with my team of computer scientists, a media scholar (me), and an artist. They include inexact computing; tiny language models; low-precision hardware architectures; hardware with limited capacity; and anticipating and mitigating energy demands at the design phase. I propose a design for an ethical, aesthetically sophisticated tiny image generator.
Looking to the economic context, I echo Bran Knowles et al.’s call for an end to digital exceptionalism, whereby ICT gets a pass on environmental issues because its products are supposedly so valuable. Proposing a true-cost accounting for the environmental impact of machine learning, I suggest that the astonishingly fast rise of ML is driven by the shareholder-capitalist framing of networked digital media.
■ Bio
Laura U. Marks works on media art and philosophy with an intercultural focus and on small-footprint media. She programs experimental media for venues around the world and is a Fellow of the Royal Society of Canada. As Grant Strate University Professor, she teaches in the School for the Contemporary Arts at Simon Fraser University in Vancouver, Canada, situated on the unceded Coast Salish territory of the Sḵwx̱wú7mesh (Squamish), Stó:lō, Səl̓ílwətaʔ (Tsleil-Waututh), and xʷməθkʷəy̓əm (Musqueam) nations. Professor Marks is also an active member of SFU Faculty for Palestine and the SFU Council on Islamophobia.
Marks was awarded a Guggenheim Fellowship for Small Files for a Small World, a book examining the carbon footprint of streaming media. She is utilizing this fellowship to sponsor international small-file workshops in twelve cities, including Dhaka, Mexico City, Cairo, Kigali, Tehran, and Guwahati (updates can be found at smallfile.ca). Her latest book, The Fold: From Your Body to the Cosmos, is published by Duke University Press.