Search papers, labs, and topics across Lattice.
This paper investigates the substantial environmental impacts of image-generating machine learning, highlighting the overwhelming carbon, water, and land footprints associated with data centers and ML devices. The authors propose innovative solutions such as inexact computing, tiny language models, and low-precision hardware architectures, while also advocating for ethical considerations in the design of ML systems. A key outcome is the development of a tiny image generator that utilizes non-scraped data, alongside a framework for true-cost accounting of ML's environmental impact, challenging the traditional efficiency-driven narrative in ICT.
The environmental cost of image-generating ML far exceeds its efficiency gains, necessitating a radical rethinking of how we design and evaluate these systems.
Machine learning is often touted to improve the efficiency of ICT, but that small gain is overwhelmed by the enormous carbon, water, and land footprints of data centers and ML-ready devices. We survey the electricity consumption of ML applications in training and inference, focusing on electricity-intensive image generation. Our team of a computer engineer, a media scholar, and an artist explore solutions including inexact computing; tiny language models; low-precision hardware architectures; hardware with limited capacity; and anticipating and mitigating energy demands at the design phase. We will sketch our work in progress of an ethical and aesthetically sophisticated tiny image generator using non-scraped data. Looking to the economic context, we will propose a true-cost accounting for the environmental impact of machine learning and suggest that the criterion of efficiency is driven by the shareholder-capitalist framing of ICT.
