Criteria for Sustainable Use of Artificial Intelligence
423.72 KB
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
The field of artificial intelligence is growing in popularity, but the impact it has on the environment has been neglected in research and in practice. The trend of deep learning models growing massively in size and computational complexity, demanding large amounts of energy and specialized infrastructure such as data centers, is not reflected in results; the performance gain is generally marginal. Moreover models are often constructed with unclear parameters or expensive methods, presenting barriers to research. This approach is known as Red AI. As such there is much room for incorporating sustainability demands and practices within the field of AI itself, known as Green AI, motivated by both efficiency and the environment.
This work aims to advance Green AI by investigating the current state of the research field via a Systematic Literature Review. The SLR addresses both the lifecycle and specific details related to building and using deep learning models more sustainably. Also discussed are tools and methods for estimating the lifecycle costs and thus environmental impact of models and the hardware they require.
Based on the contents of papers in the SLR there are industry-wide opportunities for improving the efficiency of deep learning throughout its lifecycle. Incorporating sustainability in this manner would allow for cheaper, more efficient models while lowering the barrier to participating in state-of-the-art research. However, Green AI is a complex topic with limitations and further work is needed on issues such as universally applicable metrics which consider sustainability and co-operation between AI research and industry. To help achieve this, a collection of criteria is presented in this work. The collection is based on implementing Green AI techniques to solve Red AI issues. The collection provides a novel listing of methods, principles and techniques for building more sustainable artificial intelligence and works as a starting point for those interested in the topic.