The GenerIA Blog
AFNOR SPEC 2314: Best Practices in Frugal AI
From project design to end-user acculturation, frugal AI is above all a matter of best practices. Numerous and complementary, these BPs are detailed in AFNOR SPEC 2314. Here is a thematic summary.
Note: SPEC 2314 has not been translated into English yet. The translation of the Best Practices titles below is therefore GenerIA's and may differ in the final, official English version of the document.
#1 Continuous Assessment: before, during and after
Before developing an AI system, it is essential to define the organization's real needs, based on a few key questions. What problem does the envisioned AI need to solve? What are the objectives to be achieved? With what technical and energy resources?
A rigorous upstream assessment helps to avoid the development of superfluous solutions, and to focus resources on high value-added projects. Once the AI system has been deployed, it is equally important to put in place mechanisms to measure the effectiveness of the solution, identify areas for improvement and adjust practices accordingly.
This monitoring results from two parallel strategies:
- Continuous metrology: implement tests to measure the effectiveness and efficiency of models in production.
- Continuous improvement: adopt an iterative approach to fine-tune models based on feedback.
- Relevant AFNOR SPEC 2314 best practices:
- BP 01 - Use needs analysis methods to implement frugality
- BP 13 - Manage the environmental performance of AI systems
- BP 23 - Estimate model consumption beforehand
- BP 25 - Evolve measurement strategies according to issues and constraints to maintain AI service frugality
- BP 31 - A/B test models to identify the one with the best performance/resource ratio
#2 Data selection and optimization
Data quality determines the success of an AI project. AFNOR recommends giving priority to data sets that have been established as relevant, diverse and representative. It is also important to ensure that the data selected is accessible and usable without requiring excessive resources. In this respect, the use of open data and collaborative sources contributes to a more frugal approach.
To check this mark, two preparatory steps are essential:
- Targeted data collection: collect only the data required to meet a specific need.
- Data pre-processing: clean and transform data to maximize quality and relevance.
- Relevant AFNOR SPEC 2314 best practices:
- BP 06 - Control data volume
- BP 07 - Work on data quality
- BP 08 - Use a relevant dataset to design the AI service
- BP 10 - Define data storage rules according to use
- BP 15 - Compress data
- BP 19 - Use open-source datasets for prototyping
#3 Algorithm selection and optimization
Algorithm optimization is a key step in achieving simple, efficient models that guarantee frugal AI. This optimization involves selecting less resource-intensive algorithms and/or adapting existing algorithms to make them more efficient. It can also include porting certain calculations to languages closer to hardware (C, Rust, etc.).
AFNOR also encourages the use of techniques such as pruning and quantization (adoption of lighter numerical data types), which reduce the size of models while maintaining their performance and increasing their explainability.
Two compasses should therefore be followed:
- Simplicity: less complex algorithms are easier to interpret and deploy.
- Transparency: decisions taken by more simple models are more easily explained, thereby boosting user confidence.
- Relevant AFNOR SPEC 2314 best practices:
- BP 02 - Select and develop the solution to meet the specific need, considering alternatives to AI
- BP 03 - Use compression methods to reduce the footprint of AI algorithms
- BP 04 - Define criteria to justify model retraining
- BP 26 - Write code that can be improved by several people and re-implemented on several environments
- BP 27 - Rationalize models
- BP 28 - Decompose a large AI model into several smaller models
- BP 29 - Reuse trained algorithms and share them (open-source)
- BP 30 - Favor more frugal base models
#4 Energy efficiency of infrastructures
AFNOR rightly emphasizes the importance of the choice of computing infrastructures in the energy efficiency of AI models. Indeed, in addition to the electrical specifications of the servers and processors needed for training and production, the energy efficiency (PUE) of datacenters and the specific features of the energy sources that power them (hydroelectricity, nuclear, coal...) are key factors in the environmental impact of AI. Taking them into account right from the design phase can help make deployments more virtuous.
Two main criteria are therefore to be considered:
- Energy efficiency: favor infrastructures (data centers, hardware, etc.) and algorithms that consume less energy.
- Sustainability: design the lifecycle of AI projects based on sustainable practices, starting with the choice of electricity source and suppliers.
- Relevant AFNOR SPEC 2314 best practices:
- BP 20 - Optimize the use of existing equipment
- BP 22 - Favor existing user/employee terminals for AI service training or inference
- BP 24 - Ensure infrastructure frugality throughout operation
#5 Collaboration and sharing of standards, knowledge and measurement tools
Collaboration between the various stakeholders in the sector is essential to promote frugal AI. AFNOR encourages the dissemination and sharing of best practices, tools and resources. Partnerships between client organizations, AI makers and Academia foster innovation and minimize both costs and duplication of effort.
Such a collaboration is based on two inseparable pillars:
- Knowledge sharing: exchange best practices and lessons learned within the community.
- Common standards: adopt standards that promote interoperability and compatibility between systems.
- Relevant AFNOR SPEC 2314 best practices:
- BP 09 - Integrate frugality into AI relevance criteria
- BP 12 - Set up a governing body to question and assess frugality
- BP 16 - Identify & mobilize a pool of frugal AI skills
- BP 17 - Set up and manage a single repository of frugal AI services
- BP 18 - Offer off-the-shelf AI digital products that promote frugality
- BP 21 - Create a repository for the environmental impacts of projects
#6 Ethics and responsibility: awareness-raising and training
For best practices of frugal AI to be adopted, it is essential to raise awareness and train the teams involved in the development and use of the solutions. AFNOR proposes themes for training programs to help professionals understand the challenges of frugal AI and integrate these principles into their daily work.
Among the issues involved, two of them require particular attention:
- Bias and fairness: biases will exist in data and models and must be minimized if not eliminated.
- Social impact: assess the social impact of the AI solution under consideration, to ensure that it benefits the greatest number of people, without negative side effects.
- Relevant AFNOR SPEC 2314 best practices:
- BP 05 - Implement eco-design measures in the development phase
- BP 11 - Provide for end-of-life in the management of an AI project
- BP 14 - Acculturate and train stakeholders
Conclusion
Validation of needs, energy efficiency of infrastructures, optimization of data and algorithms, acculturation of developers and users... the best practices presented in detail in the AFNOR framework promise AI solutions that are at the same time effective, sustainable and ethical. Because they lead us to rethink advanced technologies, could some of them one day become binding regulations?
In the GenerIA blog:
