Artificial intelligence is in seemingly everything these days, and if it’s not there yet, advocates say it will be soon. But training those AI models that power business processes, financial services and cities takes massive amounts of energy – even though AI is often used to monitor energy consumption to save costs and emissions.
Globally, the push towards greater sustainability and environmental conscientiousness has meant companies need to better evaluate processes and tools to accurately understand their own carbon footprint. But understanding AI’s footprint is in its infancy.
“Companies are just now at the start of thinking of AI as part of the [sustainability] equation –and it’s not just heating and cooling,” Simone Larsson, an AI evangelist at Dataiku, said during an interview with ITP.net at a conference in London.
Training even common AI models can emit more than 626,000 pounds of carbon dioxide equivalent, or nearly five times the lifetime emissions of the average American car – including its manufacturing, one study from the University of Massachusetts Amherst found.
“Large organisations want to do this ‘everyday AI’ at scale and are also grappling with the tension of CO2 mandates and making sure they’re within the threshold. They want to get to net zero, but they want to do AI,” Larsson said. “So how do they get the balance right?”
Frugal AI may be the answer. Frugal AI can help companies that only have access to small datasets train models, but it can also help companies with large datasets reign in AI scaling costs, lower their carbon footprint and align internal mandates.
Little seems to have been written on the topic so far, at least in the public domain (a Google search returns few results), but Larsson advocated for such an approach.
“It’s all about using smaller data sets, and I think it’s reusing models so you have to train the model less,” she said. “It’s using less data to train mundane operational models, and when it’s R&D [research and development] – that’s when you pull out the mammoth data set.”
Large data sets are generally assumed to create better, more accurate models.
In a blog series on the topic, Larsson addressed that training on small datasets seems counterintuitive to existing AI and machine learning norms. At the same time, over collecting data has become the norm as well.
“It has become more affordable to train and test machine learning models, and it is estimated that, over the last 10 years, there has been a 30 percent year-on-year decline in compute costs. What if the decline in costs, the desire to collect everything, and lowered barriers to entry for AI leads to habits that are unsustainable and adversely impact the environment?”
For those looking to make their AI models greener, different models of AI frugality exist. Input frugality, learning process frugality and model frugality are three options laid out in a paper titled Frugal Machine Learning.
Frugal inputs, which emphasis the associated cost of the data, may involve fewer training data or fewer features. On the other hand, learning process frugality emphasises the cost associated with the actual learning process and the computational and memory resources.
“Frugal learning might produce a model with lower prediction quality than achievable in a non-frugal setting, but do so much more efficiently,” the paper’s authors wrote.
Finally, model frugality focuses on the cost of storing or using a machine learning model. These models may require less memory and produce predictions with less computational effort compared to other models.
There are a few models outlined, but the notion of frugal AI is far from becoming mainstream. Further, there are contradictions within the frugal AI that need to be worked out yet, meaning it is likely to be some time before organisation widely adopt and put it into practice.
“Just the act of collecting all the data points for a company to baseline where they are in terms of their carbon footprint, that creates a lot of data,” Larsson said.
