Announcing the winners of the Frugal AI Challenge 🌱
In partnership with the AI Action Summit, HuggingFace, Data for Good France and the French Ministry of Ecological Transition the Frugal AI Challenge promotes the development of energy-efficient AI models addressing critical climate issues. Over the past month, participants from academia and industry were invited to submit models that not only delivered high performance but also minimized energy consumption, thereby encouraging sustainable AI practices.
Today, we are announcing the winners!!
Why Frugality ?
In this challenge, we wanted to switch the narrative we often see in typical Machine Learning competitions and more generally in the AI industry : only aiming for performance. So we developed a challenge open to all to develop innovative approaches for energy-efficient algorithms specialized to particular use cases.
So why is Frugality important ?
- Resources are constrained globally (human pressure on raw materials / energy / water) and locally (deployment in specific environments).
- We need to make sure that AI really brings net benefit (what it can enable to solve/reduce vs what it costs)
- Change the narrative - There is a current emphasis on “bigger is better” in AI - that bigger models, with more parameters, perform better.
Challenge Overview 🌎
The competition featured three distinct tasks, each targeting a pressing environmental concern:
- Text - Detecting Climate Disinformation (sponsored by QuotaClimat): Participants developed models to identify misleading or false information related to climate change in textual data. This task addressed the growing issue of climate misinformation and its impact on public perception.
🌱 In this task frugality is a requirement, because the sponsor NGO QuotaClimat needs to scale this algorithm across thousand of hours of TV and Radio transcription to measure disinformation at scale to be able to provide data to regulators. But scaling NLP algorithms with closed-source APIs can take hundreds of hours and cost thousands of $. A small efficient models could allow this kind of analysis.
- Image - Classifying Regions at Risk of Wildfires (Sponsored by Pyronear): Utilizing onsite cameras images, teams created models to assess and classify areas susceptible to wildfires and locate early signs of smoke. Given the increasing frequency and severity of wildfires globally, this task aimed to enhance early detection and prevention strategies.
🌱 For this task, frugality is neither an option, because the model needs to run on a small Rasberry Pi computer onsite in Forests in areas not connected to the internet.
- Audio - Detecting Illegal Deforestation (Sponsored by Rainforest Connection): Participants analyzed bio-acoustic data recorded in forested areas to detect sounds indicative of unauthorized logging activities. This task focused on preserving biodiversity and combating illegal deforestation practices.
🌱For this task as well, frugality is neither an option, because the model needs to run also in areas not connected to the internet.
Over the 4 weeks of the challenge in January 2025, we have received great interest in the challenge with hundreds of participants in the launch events. And we received a total of 64 final submissions across the 3 tasks.
You can look at the Frugal AI Challenge homepage and the HF Hub page if you want to browse the datasets and get your hands on developing a model.
Evaluation Criteria 📏
We received a total of 64 submissions across the 3 tasks, and they were assessed based on two primary criteria:
Performance: The accuracy of the model in completing the assigned task, evaluated based on a hidden test set.
Energy Efficiency: The amount of energy consumed during model training and inference phases, measured using Code Carbon, an open-source package for measuring the environmental impacts of code.
By emphasizing both metrics, the challenge encouraged the development of AI solutions that are not only effective but also environmentally sustainable.
Announcing the Winners 🏆
We are thrilled to announce the winners of the 2025 Frugal AI Challenge:
Detecting Climate Misinformation (Text) : "A lightweight SentenceBERT + MLP approach to lower emissions"
Classifying Regions at Risk of Wildfires (Image): “Consulting for Good"
Detecting Illegal Deforestation (Audio): "The Quefrency Guardian"
Each winning team demonstrated a remarkable balance between model performance and energy efficiency, embodying the core objectives of the Frugal AI Challenge.
For more information about the winning approaches, check out the presentation of winners from the AI Action Summit Sustainability side event!
Looking ahead 🚀
The success of the 2025 Frugal AI Challenge shows the potential for AI to contribute to environmental sustainability when developed with a specific attention to energy consumption and frugality. We extend our heartfelt congratulations to the winners and gratitude to all participants for their innovative contributions.
We will put all the datasets and evaluation criteria up on the Hugging Face Hub, and hope that these will continue to be used by others in hackathons and projects.
We encourage the AI community to continue prioritizing frugality in model development, ensuring that technological advancements align with sustainable values.