Infinite Analytics

Stories

Carbon Footprints of AI & How to Manage Them

by Team IA, March 22, 2021

AI tools never cease to amaze whether they help us identify our favourite songs by recognizing and matching humming or whistling sounds or whether they help in devising better methodologies of vaccine distribution, or even in making mental healthcare more accessible. It is an oxymoron then that AI which helps humankind in making accurate climate change predictions and smart grid design etc., itself has a huge carbon footprint problem which needs to be addressed on war footing.

University of Massachusetts, Amherst, researchers have found that  training an AI model leads to CO2 emissions of up to 284 tonnes which is equal to five times the lifetimes emissions of an average car in America, including the car’s manufacturing process as well! Even as natural-language processing (NLP) models are making continuous strides and impressing with their sentence completion, dialogue and conversation, machine translation etc, its model training process requires training on big data sets. Not only is this expensive but the method is substantially energy intensive. In their research, it was also found that NAS OR neural architecture search, including the automation of a neural network’ design through trial and error – was highly energy intensive.

Shrinking the carbon footprint | Steps to tackle the situation at hand

Agreeing on paper to resolve the issue

Tech giants such as Amazon and Google are investing in renewable energy to reduce AI training carbon emissions. In fact, September 2019 saw employees of Amazon, Microsoft, Google, Facebook, and Twitter, join a worldwide march against climate change whereby they demanded that their employers should issue an assurance towards reducing emissions to zero by 2030. This can be made possible when the tech giants enter into deals with fuel companies coupled with putting an end to the exploitation of climate refugees. Tech giants can also lead the way for others to follow in reducing emissions.

Hardware for better efficiency of deep net algorithms| Efficient deep nets

Researchers are looking for alternatives such as optical computers that use photons in place of electrons and quantum computers which are capable of increasing the computing power etc. To decrease the carbon footprint, researchers are looking at finding ways of storing and reusing data locally instead of shuttling data from a dedicated memory site. This speeds up model training and makes deep learning apps to run even more efficiently.

Marching towards Green AI

Allen Institute of Artificial Intelligence is credited with starting the Green AI movement through their research work. Their paper roots for AI research yielding the desired results without making the computational cost any higher, or even reducing it in some processes.

Besides these measures, AI companies could also look at carbon transparency, accounting for the full-stack supply chain, saying no to fossil-fuel consumption, tech regulations and green deal policy making should be integrated, and more importantly highlighting and knowing what and where AI harms and its impact on climate refugees.

Subscribe to our newsletter to keep updated about the world of AI and Infinite Analytics!

Subscribe to our newsletter

* indicates required



   


Avatar

by Team IA

Leave a comment

Your email address will not be published. Required fields are marked *