Harnessing the potential of artificial intelligence (AI) while tackling its environmental drawbacks, MIT researchers introduce an innovative ‘once-for-all’ (OFA) network, designed to minimise carbon emissions and transform the AI environment.
You would have to be living under a rock for the last year or so to not know that Artificial Intelligence (AI) is a technological big deal. A marvel with enormous strategic potential and transformative power, the exponential growth of the AI environment seems set to influence us more and more in our daily lives. Perhaps that’s why the only downsides that have caught the public imagination so far relate to fears about the technological singularity – we worry about the dystopian cliché of being assimilated by the Borg. But there’s something else and it’s more science fact than fiction.
With great power comes environmental impact
AI also presents significant challenges with its high energy usage, leading to environmental concerns. Enter a pioneering team of researchers from the Massachusetts Institute of Technology (MIT), who have developed a solution that not only minimizes costs but also critically diminishes the carbon footprint associated with training AI models.
In 2019, a shocking revelation was made by the University of Massachusetts at Amherst about the colossal energy consumption of AI model training. It was found that this process released an astounding 626,000 pounds of carbon dioxide, equating to the carbon emissions from an average American car, including its manufacture, amplified fivefold. This is largely because contemporary AI operates not merely on personal computers or straightforward servers, but deep neural networks are implemented on an extensive assortment of specialized hardware platforms, intensifying energy consumption.
This trend of high energy use is further evidenced by Google’s AlphaGo Zero, an AI system that plays Go against itself to learn independently. Analytics Insight and Kepler Lounge highlighted that this AI model produced an overwhelming 96 tons of carbon dioxide over just 40 days of research training. To put this into perspective, this emission is equivalent to a thousand hours of air travel and the yearly carbon footprint of 23 American households. The implications are clear: maintaining such rates of AI deployment could result in unfeasible and unsustainable conditions.
The research team at MIT, however, has forged a path towards an environmentally conscious AI environment, pioneering an automated AI system named the ‘once-for-all’ (OFA) network. This novel system is based on the advancements of automatic machine learning (AutoML), and it targets the root issue of energy consumption by decoupling training and search processes to reduce cost.
In essence, the OFA network serves as the ‘mother’ network to a series of subnetworks. As the parent network, it imparts its knowledge and previous experiences to the subnetworks, effectively training them to operate autonomously without the need for continued retraining. This method stands in stark contrast to previous AI technologies, which necessitated repeating the network design process and retraining from scratch for each deployment scenario, leading to a linear growth in cost, excessive energy consumption, and significant carbon dioxide emissions.
With the OFA network in operation, the need for repetitive retraining of subnetworks is practically eliminated. This efficiency drives down costs, slashes carbon emissions, and enhances overall sustainability, creating a more environmentally friendly AI environment.
Assistant Professor Song Han from MIT’s Department of Electrical Engineering and Computer Science, who led the groundbreaking project, commented on the significant environmental impact of their work. He stated, “Previously, searching for efficient neural network architectures resulted in a massive carbon footprint. However, we have managed to reduce that footprint exponentially with these new methods.”
Chuang Gan, the co-author of the MIT research paper, also expressed his enthusiasm about the compact nature of the model. He stated, “OFA is pushing the boundary of efficient deep learning on edge devices due to its compactness.”
This compactness could mean that the future of AI lies in miniaturization, opening up new opportunities for environmentally friendly operations. By significantly reducing the energy requirements and carbon emissions of AI model training, this innovative approach by MIT offers promising steps towards a sustainable AI environment.
