AI has the potential to help us tackle the problems associated with climate change and the warming of our Earth. The closer we get to the precipice, the greater the urgency. This has helped fuel tremendous growth in AI projects throughout government and the public sector, where AI is being used to make more accurate climate change predictions or to intelligently power the infrastructure that could support lower emissions on a global scale.
Amidst all this enthusiasm, the one thing often being left out of the conversation is the carbon cost of these compute-intensive solutions. At best, the adoption of AI might be slowed down because people haven't adequately considered the cost (financial or environmental) of the solution required. At worst, it could accelerate the warming of our planet.
This is why it is so important to develop a Green AI technology: a technology that takes into account energy-efficiency as an important evaluation metric.
The world is rapidly changing
97% of scientists agree that humans are one of the main drivers of climate change. The Paris Agreement attempts to hold governments accountable for their role in the crisis and sets a goal of limiting global warming to 1.5ºC, compared to pre-industrial levels. According to the European Union's statements on “Supporting the green transition” [European Union Report, Feb 2020], today, the information and communication technology (ICT) sector accounts for 5-9% of all electricity use. This portion equals more than 2% of global greenhouse gas emissions and roughly the same as all air traffic in a normal, non-pandemic year. As AI gets integrated into more and more aspects of society, including smart cities and driverless cars, this portion will climb. Some see the ICT share of all electricity use increasing to 14% by 2040. The question is, what role can and should AI play in this?
There’s no free lunch.
It has been clear for some time that AI’s carbon footprint could eventually come to dirty the heralded fourth industrial revolution. AI growth currently comes at a cost: the algorithms behind AI solutions are computationally expensive to train and deploy, which results in significant barriers to entry for smaller businesses and individuals and the creation of a large AI carbon footprint.
An alarming trend in published research for AI gives a glimpse of where this is headed. In 2018, 75% - 90% of publications focused on AI solutions that favoured accuracy over efficiency. When put into practice, the large machine learning models that emerge from this type of research consume massive amounts of energy. It has been estimated that the carbon footprint created by training one of these models is equivalent to 300,000 kg of carbon dioxide, which is the same as making 125 round-trip flights between New York and Beijing.
The debate around the impact of digital technologies on the planet Earth has emerged in recent years and has been reported by many prominent institutions and governments. The recent report from The Royal Society on “Digital Technology and the Planet”, highlights that:
“Data-enabled technologies such as machine learning and artificial intelligence have also enabled efficiencies and optimisation across sectors. While some of these applications already contribute to reductions in greenhouse gas emissions, digital technologies also have an environmental cost which should not be neglected – from the extraction of minerals to the energy use and emissions of the technology”. [Royal Society, 2020, p 6]
One example above all is the blockchain technology associated with mining cryptocurrency, used to securely transfer, verify and record digital asset ownership. This technology involves heavy computations and a consequent high carbon footprint, which has been growing incredibly fast in the past years. It made the headlines in February 2021 when Bitcoin’s carbon consumption surpassed Argentina’s as a nation (BBC), according to the Cambridge Bitcoin electricity consumption index CBECI.
Experts agree that the use of digital technologies and AI solutions can help reduce carbon emissions across industries. However, a too-high computational cost of these technologies could slow down their adoption, which can impact the effort to combat climate change.
What is Green AI?
Green AI is an emerging topic in the AI community, which started to gain popularity only recently. The early work of 2019 from Schwartz et al. at the Allen Institute for AI “...advocates a practical solution by making efficiency an evaluation criterion for research alongside accuracy and related measures”. This approach, at the core of the Green AI philosophy, will help to “make AI both greener and more inclusive”.
At Mind Foundry, we believe that making AI digital transformations sustainable can only be achieved by humans and AI working together. This is why we are addressing the market need to enable these workflows in our core product offerings via the Green AI Auditor. Mind Foundry is committed to enabling humans and AI to work together to address the problems that impact people’s lives at scale and enact positive change. The Green AI Auditor will bring to our platform the capability to enable our users to predict, monitor and mitigate their carbon footprint generated through the computational requirements of their AI technology projects. Users will be made aware of the costs, both financial and environmental, of their model building and model deployment projects and provide them with tools for managing the resulting trade-off between how accurate their model’s predictions are and the amount of resources required to generate that accuracy.
Learn more about Green AI in this 4 minute video presentation
The business use case: Reducing resource wastage when developing solutions
Building a bespoke AI solution is often an iterative process where the data scientist wishes to quickly assess many different approaches and develop only the most promising solutions. Determining an efficient strategy for assessing each solution is challenging for various reasons, related both to the data and the resources sides of the problem.
It is difficult to know in advance how much data is required to assess a solution. Too much data can lead to very expensive calculations, and too little data, on the other side, can lead to inconclusive answers. On top of this, access to big-data technologies has become more widespread, and many tools make it easier to solve difficult problems merely by using more computation and, therefore, energy. There is often little consideration given to whether equally good solutions can be reached using less data and, hence, less computation and energy.
A key part of this process is incorporating a cost-benefit analysis into your system and accurately assessing what you're compromising. For example, if you can reach an equally good solution with less computation, then surely that's better. But what happens if you can make a solution that's 2% more accurate but 20% less efficient? Is that an appropriate tradeoff for you? How does that stack up against your environmental obligations? Or your performance obligations? It's going to be different for each application and each industry. With the Green AI Auditor, Mind Foundry is providing the ability for businesses and data scientists to make informed decisions about these important business-critical decisions.
It is also difficult to allocate in advance an adequate amount of resources (time, memory etc) to perform a given calculation. The resource budgets are often not planned in advance, and it can happen that an experiment is launched and then fails because it hits memory limits. At this point, the calculation is abandoned, resulting in a waste of resources (electricity and time).
The data scientists in charge of building a new AI solution have to face all these challenges, and we want to empower them with the ability to focus on how to solve the problem and not on the practicalities of running the solution, including budgeting for limited memory, time or other resources.
Developing new technologies
To solve the problem described above, the Mind Foundry Green AI auditor develops several new capabilities within our platform, falling into two main categories:
Oversight: predicting and monitoring resource and carbon consumption during AI model building and deployment
Mitigation: managing the exploration of, or the optimisation of, a position on the accuracy-efficiency trade-off.
With these capabilities, users of our platform will be both aware of the carbon and financial costs associated with developing and deploying their AI systems and have tools to help reduce these costs when desired.
Our users will be explicitly made aware of the costs, both financial and environmental, of their model-building and model-deployment projects, and they will be equipped with the tools for managing the resulting trade-off between model accuracy and model efficiency.
The technology underpinning the Green AI solution includes a cost-aware Bayesian optimiser, which leverages a Bayesian approach to sequential decision-making and allows the user to manage the accuracy-cost trade-off.
A cleaner world
The digital revolution is coming at a cost for the environment, with an increase in CO2 emissions caused by the digital sector, including AI. We believe this is the best way to unlock the potential for AI to be a key part of the solution to the climate crisis and not an unwitting cause.