I will be referring to the following 3 papers on this very interesting topic.
(1} https://link.springer.com/article/10.1007/s43681-021-00043-6
Sustainable AI: AI for sustainability and the sustainability of AI
A Van Wynsberghe – AI and Ethics, 2021 – Springe
(3) Lacoste, A., Luccioni, A., Schmidt, V., Dandres T.: Quantifying
the Carbon Emissions of Machine Learning. (2019)
While there is a tremendous push for using new-generation generative AI based on large language models to solve business applications, there are also voices of concern from experts in the community about the dangers and ethical consequences. A lot has been written about this but one aspect which has not picked up sufficient traction, in my opinion, is Sustainable AI.
In (1), Wynsberghe defines two disciplines on AI & sustainability. AI for Sustainability and Sustainable AI.
AI for Sustainability is any business application using AIML technology to solve climate problems. Use of this new generation technology to help in climate change and CO2 reductions. Major applications are getting developed for optimal energy distribution across renewable and fossil energy sources. Any % extra use from renewable sources, help in less use of fossil fuels and help in climate change. Various other applications may include better climate predictions and the use of less water, pesticides, and fertilizers for food production. Many Industry 4.0 applications to build new smart factories, smart cities, and smart buildings fall into this category.
On the other hand, Sustainable AI measures the massive use of GPU and other computing, storage, and communications energy usage while building the AI models and suggest ways to reduce this. While digital software development and testing can be done in a few developers’ laptops with minimal use of IT resources, the AIML software development life cycle calls for the use of massive training data and develop deep learning neural networks with multiple millions of nodes. Some of the new generation Large Language models use billions of parameters beyond the imagination of all of us. The energy use does not stop here. Fine Tuning learning for specific domains or relearning is as energy-consuming or sometimes even more than original Training. Some numbers mentioned in (1) are reproduced here to highlight the point. One deep-learning NLP model consumed energy equivalent to 600,000 lbs of CO2. Google Alpha-Go-Zero generated over 90 Tonnes of CO2 over 40 days it took for the initial training. These numbers are large and at least call for review and discussions. I hope I have been able to open your eyes and generate some interest in this new dimension of AI & Ethics i.e impact on climate change.
I am sure many of you will ask “Isn’t any next-generation industrialization from horse carriages to automobiles or steam engines to combustion always increased the use of energy and why do we need to worry about this for AI?”. Or “There has been so much talk on how many light bulbs one can light for the same power used for a simple google search , why worry about this now ?”. All valid questions.
However, I will argue that
- The current climate change situation is already in a critical stage and any unplanned large-scale usage new of energy can become “the feather that broke the camel’s back!”.
- Use of fully data driven life cycle and billions of parameters, deep neural networks are being used for the first time at an industrial scale and industry-wide and there are too many unknowns.
What are the suggestions?
- Energy consumption measurement and publication must become part of the AI & Ethics practice followed by all AI development organizations. (2) Carbon Tracker Tool and (3) Machine learning emission calculator are suggestions for this crucial measurement. I strongly recommend organizations use their Quality & Metrics departments to research and agree on a measurement acceptable to all within each organization. More research and discussions need to calculate the net increased use of energy compared to current IT tools to get the right measurement. In some cases, the current IT tools may be using legacy mainframes and expensive dedicated communication lines using up large amounts of energy and the net difference by using AIML may not be that large.
- Predicting the energy use at the beginning of the AIML project life cycle also is required. (3).
- The prediction data of CO2 equivalent emissions need to be used as another cost in approving AIML projects.
- Emission prediction also will force AIML developers to select the right size training data and use of right models for the application. Avoid the temptation of running the model on billions of data sets just because data is available!. Use the right tools for the right job. You don’t need a tank to kill an ant!.
- Ask the question of whether the use of deep learning is appropriate for this business application? For example, a simple HR application used for recruitment or employee loyalty prediction with Deep learning models may turn out to be too expensive in terms of Co2 emissions and need not be considered a viable project.
- CEOs include this data in their Climate Change Initiatives Report to the Board and shareholders and also deduct carbon credits used up by these AIML applications in the company’s Carbon credit commitments.
More Later,
L Ravichandran
