In recent years, the evolution of Artificial Intelligence (AI) has experienced super-linear growth. This growth is holistic, as it is evident across various aspects of the entire lifecycle of AI, such as data collection, model optimization, infrastructure, and hardware production and selection. However, the exponential scaling of AI and the pursuit of higher model quality have led to a significant increase in its carbon footprint. To give some context, research showed that training a large Machine Learning (ML) model like Meena from Google has a carbon footprint equivalent to driving 242,231 miles, while ML training at Meta averages 1.8 times that. That’s like a round trip to the Moon and back. This is proof that this is a pressing issue that demands urgent attention and sustainable solutions in AI development. My proposed solution to this problem is a web-based application that allows users to input details about their AI systems (e.g., dataset size, hardware type, runtime, model selection) to estimate its lifecycle environmental impact in terms of carbon footprint. The key features will include an intuitive user interface, real-time feedback, and comprehensive tracking of carbon emissions throughout the development process. It will also have a built-in recommendation system that would suggest alternative methods if the details of the user input point to a significant environmental impact. The poster will further detail the application's functionality, key features, and its potential to significantly reduce the environmental impact of AI development including specific examples, use cases, and potential future enhancements.
Temitope Adeyelu (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: