"Green" Artificial Intelligence

25

English

ESPCI -PSL (to be confirmed)

This formation aims to provide the context and knowledge necessary to optimize the development and deployment of AI-based solutions while considering environmental issues. ;
The objectives of the course are as follows:

- Understand the environmental challenges of numerical solutions and quantify their various ecological impacts.
- Explain general principles of numerical optimization for scientific computing, then explore optimization aspects specific to AI models.
- Question the utility and environmental implications of AI model deployment.
- Discuss the possible societal benefits of AI and the ethical issues that arise.

The course spans multiple days, incorporating theoretical presentations, practical exercises, and expert interventions.
* The following program is temporary and subject to changes. *;

Part I. (1-2 days);
*** Environmental challenges of numerical methods ***;
- Reasons and urgency to reduce the environmental impact of numerical technologies.
- The different types of environmental impact and the proper metrics to quantify them.
- The different types of numerical usage and their specific impacts.
- Estimate and measure the energy consumption of AI solutions.

Part II. (2-3 days) ;
*** Optimization of general scientific computing and IA model specificities ***;
- Optimization principles related to time, costs, and resources.
- Impact of programming languages, libraries, and hardware choices on environmental efficiency.
- Computationally intensive operations in AI models and optimization.
- Impact of AI model architectures. Construct metrics that consider the balance between predictive strength and numerical efficiency.
- Utility of AI model deployment versus lighter alternatives. Use of lightweight AI models to accelerate existing heavy computations.

Part III. (1 day) ;
*** How does AI impact society? How to settle the benefit / environmental-cost ratio? ***;
- Domains where AI can significantly benefit society regarding decision support, process optimization, etc.
- Ethical and environmental problems related to AI usage, especially in polluting industries or cases involving restricted freedoms.
- Discuss control mechanisms, social acceptability, and branding of AI.
- Present concrete cases of beneficial AI usages and some existing "Green AI" projects or initiatives.

The week's validation is only subject to full-time participation and the production of some performance versus consumption measurements during the corresponding practical work sessions.

This course is mainly intended for frequent users or developers of AI models in a scientific context.
We strongly recommend that the participants are already familiar with the basics of artificial neural networks.
Still, any person interested in numerical optimization and environmental issues should be able to follow the week.

- Medium proficiency in Python is required.
- Basic knowledge in at least one compiled language (C, C++, Fortran) is strongly advised.
- Experience with Keras/TensorFlow would be beneficial.

CORNU David

CORNU David (Observatoire de Paris - PSL);
ALLAUZEN Alexandre (ESPCI - PSL);