Artificial Intelligence Contributes to Climate Change. Here’s How.
AI will have a huge carbon footprint, so two students have developed a tool to combat this issue.
Artificial intelligence, machine learning, and deep learning. To many, these are just modern buzzwords synonymous with robots, computers, and programs, but they’re so much more than that. AI is transforming the way we communicate, saving lives, and making many of our daily tasks, from typing text messages to finding good shows to watch on Netflix, much easier.
Another word tossed around nowadays is climate change, which refers to the changes in global climate that have occurred in the past couple of decades due to a massive increase in atmospheric carbon dioxide. You may be wondering why I’m bringing this term up because it has nothing to do with AI, right? Wrong. AI, especially the field of deep learning, is predicted to increase atmospheric CO₂ levels and become a major contributor to climate change.
To spread awareness about this fairly unknown problem and combat it, two students from the University of Copenhagen’s Department of Computer Science, Lasse F. Wolff Anthony and Benjamin Kanding, along with Assistant Professor Raghavendra Selvan, developed a program that predicts the CO₂ emissions of training deep learning models. Let’s take a look at the relationship between AI and climate change and see where this program fits in.
AI and Climate Change: A Complex Relationship
Many people believe that AI will be used to combat climate change, and they’re not wrong. From figuring out ways to optimize energy efficiency to creating better estimates of power usage, machine learning has several opportunities for helping to reduce the effects of climate change. This is great and all, but there is one major downside: it requires lots of energy.
Deep learning (DL), a subset of machine learning, focuses specifically on the use of artificial neural networks to do tasks like classification and generation. DL models are extremely powerful, but the energy required to train them has grown 300,000-fold (from 2012 to 2018)! More powerful DL models must be trained for hundreds or even thousands of hours on powerful computers, and to meet this high energy demand, energy production must be increased. In 2010, energy production released 35% of the total greenhouse gas emissions, and this number is sure to increase as DL gets more popular.
“Developments in this field are going insanely fast and deep learning models are constantly becoming larger in scale and more advanced. Right now, there is exponential growth. And that means an increasing energy consumption that most people seem not to think about,” according to Anthony.
One popular deep learning model, GPT-3, uses the same amount of energy as 126 Danish homes do in a year and emits the same amount of CO₂ as over 430,000 miles of driving, all in one training session! That may sound scary, but models will get even powerful in the future, which means even more energy will be consumed.
Kanding says, “As datasets grow larger by the day, the problems that algorithms need to solve become more and more complex.”
What can we do to mitigate AI’s effect on the climate? Anthony and Kanding believe they have come up with a way to spread the word about AI’s negative impact and help combat it.
Carbontracker: A Solution to AI’s Carbon Footprint
As I previously stated, two students from the University of Copenhagen, Lasse F. Wolff Anthony and Benjamin Kanding, have created carbontracker, a tool that predicts the energy consumption and carbon emissions of training a deep learning model. To create it, the students used Python, which is used extensively for DL, for trying to predict the total training time and carbon footprint of a model.
Carbontracker does not directly reduce the carbon footprint of training a DL model, but it lets the programmer/researcher be aware of the negative impact that training their model is going to have on the environment. This allows them to try to make their model less power-hungry or even contemplate whether or not they want to train the model.
Additionally, in their research paper, Kanding and Anthony have provided several ways to reduce an AI’s carbon footprint. Solutions include:
- Training your model in places where less carbon is emitted from electricity production (ex. Sweden, Estonia)
- Training your model whenever carbon intensity is lowest during the day (less carbon may be emitted during certain hours of the day)
- Choosing more efficient algorithms, which can reduce the computing power needed for training
- Using more efficient hardware (ex. using more energy-efficient GPUs)
It seems like it was just recently that artificial intelligence was ‘something of the future,’ but it is rapidly gaining popularity now. Unfortunately, it negatively impacts our already worsening climate, and this is a problem we can’t just ignore. Our world runs on energy, but energy is contributing to the changing of our world’s climate (how ironic), and we need to take any step we can to combat this trend. Carbontracker is a good start, and hopefully, we can one day train a DL model without worrying about its carbon footprint.
To learn more about carbontracker, read the official research paper:
