The sheer size of datasets and the number of computations required to train machine learning algorithms creates a huge cloud server workload and significantly increases the carbon footprint. The European SustainML project aimed to come up with an innovative development framework that would allow AI designers to reduce the power consumption of their applications, French research institute Inria said in a press release.
A study published in Nature found that in 2019, a run-of-the-mill training model used for natural language processing emitted 300,000 kg of CO2, equivalent to 125 round-trip flights between New York and Beijing. Five years later, all sectors of society are now enthusiastically adopting deep neural networks, and as artificial intelligence grows to unprecedented scale, so does the toll on our planet.
With this in mind, the main objective of the European project SustainML is to create a framework to help AI designers consider their energy consumption when developing machine learning applications. Janin Koch, a scientist in the Inria Saclay Center Common Offsite Project Team, leverages human-computer interaction (HCI) to help his AI designers make more sustainable decisions across his ML lifecycle. I’m digging more specifically how to help. And to raise awareness of the cost-benefit trade-offs behind each of these choices. The project will start in October 2022 and will involve Inria and other stakeholders.
Quantifying the carbon impact behind machine learning models
This project involves different research areas. One of the underlying aspects is the environmental impact of the algorithm, or more precisely, the quantification of the consequences of each decision made throughout the ML lifecycle. For example, if you choose to train your ML model in a cloud facility that relies on non-fossil renewable hydroelectric power rather than in a data center powered by a coal-fired power plant, you will notice a significant difference in the amount of carbon emitted. occur.
But this is not the end of the story. According to Janin Koch, “This is a much broader issue than just choosing a clean cloud. In fact, we need to rethink what we really need.” The trend is that the more data, the more complex the model, the better the final result, which is to some extent not entirely unfounded, especially for complex problems. No, but many applications don’t necessarily need this level of accuracy or this amount of data.” So before starting an AI project, scientists ask themselves, “What do I really need?” you have to ask yourself.
Is there a more sustainable alternative that can reduce the amount of data required and run time? Instead of collecting a lot of data, can we reuse/reuse existing datasets out there? Do you have to create and train a model from scratch, or can you reuse a model already available in a code repository? Do you really need your model to run for a long time? “Ultimately, it’s not just about improving the algorithms, it’s about improving the entire application lifecycle.”
Human-centric interactive framework
In addition to raising awareness about sustainability trade-offs, the project aims to create interactive tools that enable developers to make more sustainable decisions at every stage of the development process. is. This is where Koch’s contribution comes in handy. “My research field is human-computer interaction. I am interested in how humans and systems can work together to explore new ideas. It includes both the method and how the system creates the proposal and repeats it.”
“In the context of this project, this means: What do developers know before starting a project, and how can they describe the overall goals of the system? can be very vague at times, so consider how the system can help you decide what is needed to achieve a particular goal and which approach is suitable for that To do.“
For such a tool to work, it must be able to explain to users how decisions are reached, how conclusions are reached, and how constraints are applied. “This process is actually very difficult. If an algorithm claims that one particular decision is 80% better, what does that mean for the user? That’s not how people understand things.” Instead, she suggests contextualizing the descriptions within the project goals and user processes to make these descriptions more meaningful.
The SustainML project is expected to have a major impact on the so-called “democratization of green AI”, where not only tech giants but also SMEs, private enthusiasts, NGOs and individual innovators can develop AI in a more sustainable way. will be able to develop
