Frugal AI: A Gentle Introduction to the AFNOR SPEC 2314 Framework

By The GenerIA Team  |  October 8, 2024  |  5 m.

Share this article:

Fostering innovation without hastening the attrition of natural resources. This is the rationale behind frugal artificial intelligence, whose definition, contours and practices AFNOR intends to normalize.

It is now a well-documented certainty that the development and deployment of AI poses major challenges in terms of resource consumption and environmental impact. To contain and, why not, reverse these trends, it is essential to adopt responsible and sustainable practices. In this perspective, the AFNOR (Association Française de Normalisation) SPEC 2314 standard proposal comes at just the right time. Written in partnership with industry specialists, the rigorous document sets out valuable recommendations to guide companies and their employees in the implementation and use of frugal artificial intelligence.

In the field of AI, frugality is defined as an approach aimed at designing systems that are efficient in terms of resource consumption while maintaining high-level performance. Unlike traditional AI models, which often require huge amounts of data and therefore modeling computations, frugal AI wants to optimize the use of available resources. This includes reducing energy consumption (and thus CO2 emissions and water stress), a fair assessment of data needs, and the use of technologies accessible to the average organization, not just to tech giants.

The principles of frugal AI

The AFNOR framework identifies six fundamental principles underpinning frugal AI:

1. Resource efficiency - Frugal AI favors algorithms and models that consume fewer resources, both in terms of data and computing power. This mechanically results in a significant reduction in its carbon and water footprint.
2. Sustainability - The emphasis here is on creating solutions that have minimal environmental impact throughout their lifecycle, not just during model construction. This dimension is characterized by using inference techniques that require less energy.
3. Accessibility - Frugal AI aims to make AI technologies accessible to a greater number of organizations, including small and medium-sized businesses. This means simplifying development processes and reducing associated costs.
4. Transparency and ethics - The decisions and results produced by frugal AI algorithms must be understandable and justifiable. This explainability will encourage the adoption of frugal AI by as many users as possible.
5. Adaptability - Frugal AI must be capable of adapting to different contexts and needs, to enable the flexible (i.e. efficient) use of available resources and existing technologies that have themselves been validated as frugal.
6. Collaboration - Frugality must be conceived as the result of collaboration between the various stakeholders (organizations, researchers, governments) so that best practices and knowledge can be shared, and collective innovation fostered.

Immediate *and* lasting benefits

In addition to immediate benefits in terms of environmental impact, the adoption of frugal AI offers several significant advantages, starting with the reduction of direct costs, a logical consequence of reduced resource requirements. This economic argument is expected to play a driving role in the adoption of frugal AI. The acceleration of innovation is also a major consequence of frugality: as frugal AI is accessible to the greatest number, SMEs and startups can innovate more rapidly, stimulating competition and the development of new solutions.

More flexible and therefore less technically demanding, frugal AI systems guarantee greater efficiency in a wider variety of usage contexts. This is particularly the case when the corpus of documents on which their models are based is smaller, and when inference infrastructures are not unlimited.

And of course, the dimension of Trust cannot be underestimated: the transparency and ethics cited above as intrinsic principles of frugal AI strengthen the trust of users and consumers in these technologies, which are no longer perceived as necessarily opaque.

The challenges of frugal AI

While frugal AI means numerous benefits, it is not without its challenges. First, there's the technical complexity. Designing frugal AI models requires advanced expertise. Organizations wishing to embark on the adventure themselves will have to invest in training and the development of specific skills - or call on specialist partners such as GenerIA.

Beyond HR skills, the right balance is also difficult to find, especially on a per-use case basis, between the performance of the models and the efficiency of the infrastructure required to achieve the desired results. To put it plainly, it is vital that the reduction in resources doesn't affect the quality of inferences, otherwise, no matter how frugal, the system won't be used.

Another important challenge is that the lack of clear standards for frugal AI can complicate its adoption. The AFNOR proposal emphasizes the need to develop a specific standardization in order to guide companies in implementing solutions that are both efficient and sustainable. Accordingly, because the concept of frugal AI is still unclear to many potential users, it is essential to raise awareness of its benefits across the board and encourage everyone to adopt its practices.

Endless examples of applications

The AFNOR framework is not just theoretical. It sets out to present several concrete examples of frugal AI applications in various sectors, including, for example:

• Agriculture, with the use of sensors and AI models to optimize irrigation and fertilization, in order to reduce water and fertilizer consumption.
• Healthcare, with the development of medical data analysis tools that require fewer “samples” to provide accurate diagnoses, in order to facilitate access to healthcare in remote areas.
• Transportation, with the dynamic optimization of delivery routes thanks to algorithms that can be run from delivery vehicles.
• Energy, with the real-time monitoring of gas supply pipelines to prevent network leaks...

Conclusion

Frugality is probably the only credible and sustainable response to the many challenges associated with AI as it is practiced today, and which the press, specialized or not, is constantly highlighting. By adopting frugal AI, organizations can not only reduce their costs and environmental impact, but also stimulate innovation and boost user confidence. To achieve this, it is essential for industry players to work together to standardize practices and thus promote a coherent multi-sector solution offering. This is one of the objectives of the AFNOR standard, which focuses on best practices.