Will AI Accelerate Climate Change?

There are many clear applications for Artificial Intelligence (AI) to tackle some of the most pressing issues, including climate change. It is obvious that moving forward, AI will be playing a role in increasing efficiency, reducing redundancies and improving our reactions to changes in our environment. But, these improvements also have a negative externality where the cost on the environment of running our deep learning algorithms continues to increase with more complex model use. It’s possible that our increased dependency on computer driven models will actually exacerbate our current climate issues more than they help to solve them. In this situation, there is a tradeoff to be made: Do we increase our computing power and model growth leading to extra CO2 emissions today in order to reduce emissions further down the road? Or do we begin to limit them now and limit our usage of potential climate saving AI?

First we can ask, is it possible that AI can help to prevent further climate change, and possible work to reverse existing effects of climate change? Recently, a group of researchers has focused on applications of machine learning (ML) that can help tackle growing problems that are a result of climate change and mitigate further emissions output. In the paper, they cover a wide range of areas where algorithms can provide leverage in solving this complex issue. Some of the applications they highlight as “high-leverage,” where ML is well suited to the application, include demand forecasting for energy sources, freight routing and consolidation, smart buildings, supply chains, and precision agriculture.

AI can be used to deal with and mitigate some of the consequences we already see today as a result of climate change. We can use AI to help predict flooding and damage to communities that is associated with extreme weather, allowing for better preparation. Google has used ML for flood prediction, where the researchers achieve 90% recall and 75% precision in flood extent models. This is a step in the right direction, but as the authors note there are challenges in this field due to the high computational cost of the simulations, errors in inputs and especially the need for manual calibration for new locations. Flood modeling can help us preemptively know the extent and possible damage inflicted by storms. This not only can help to reduce the economic burden of the aftermath of a flood but is also potentially life saving. While not solving the challenge of climate change, AI can help us better deal with the consequences such as extreme flooding. Startups like Jupiter Intelligence are using AI to predict risk from climate change impacts and natural disasters. Continuing to build out a library of models and data to address different types of natural disasters will be an important step moving forward. While we still have a long way to go, as seen by this, “is it a toilet or a flood?” confusion by ImageNet when looking at the aftermath of Hurricane Maria in Puerto Rico, steps are being made in the right direction.

The benefits of using AI and ML to tackle climate change are clear, but it is also important to remember the cost of developing and producing these models. As our algorithms becoming increasingly complex, there is an equivalent increase in computation power used to train and run them. 

Recently, a group of researchers published a paper tackling the carbon emission output of different ML models. Focusing on deep learning for natural language processing (NLP), the researchers note that with the increased accuracy amongst NLP tasks there has been a corresponding increase in computational resources needed to reach new benchmarks. Specifically, they find that training a state-of-the-art model transformer model with neural architecture search (NAS) can produce an estimated 656,347 lbs of CO2 emissions. To benchmark this amount, an average car including fuel has an estimated 126,000 lbs of CO2 emissions over its entire lifetime. Not only does it come with a high level of emissions, the authors also estimate a cloud computing costs of a little under $1M to about $3.2M. Throwing a lot of computational power at a model might not be worth the increase in accuracy. Researchers at AI2 are tackling the discussion of the trade off between increases in data and energy versus the marginal increases in accuracy of a model. As the authors note, the relationship between the number of experiments (neural networks) and performance growth is logarithmic.

Image: Showing the diminishing returns of increases in data from “Green AI”

We have seen some of the large tech enterprises come out with sustainability goals. Amazon has a goal of having AWS running on 100% renewable energy sources for their cloud infrastructure, with already 50% of their energy in 2018 coming from renewable sources. Google has invested in wind and solar farms around the world and are 100% powered by renewable energy sources through their energy purchases. While these types of corporations are not running solely on renewable resources yet, their purchases from the green grid spur solar and wind farm developments which moves us in the right direction. Still, it is important to highlight that the leading cloud providers need to take the lead in aggressively moving towards powering their services with renewable energy full time. 

DeepMind has already shown a clear example where AI has helped optimize energy consumption at one of the Google Data centers. By implementing AI controlled optimization, the data centers are seeing energy savings of 30 percent on average. As they also note, the savings increase over time as the AI has more data to work with and the algorithms are further improved. Implementing these energy saving measures at data centers is just one application of AI that reduces our overall emissions. This is a key example of the tradeoff of using emission heavy AI to offset future emissions. AI can help itself run more efficiently in the data centers. 

Increasing efficiency and implement renewable sources of energy are clearly steps in the right direction but there are also algorithm and model level techniques that can be implemented to reduce the energy consumption in training and using ML models. In my next post, I will highlight a couple of frameworks and techniques focused on algorithm efficiency.