Climate-friendly AI

After the course you will:

• Be able to quantify the energy consumption and carbon footprint over the lifecycle of AI models
• Be acquainted with best practices in documenting and reporting of AI climate performance
• Be able to strike an informed and appropriate balance between performance and efficiency when developing and deploying AI
• Have experience implementing a variety of techniques that can reduce the energy consumption and carbon footprint of AI, for transitioning towards Climate-friendly AI

• Motivation for Climate-friendly AI: Discuss the impact of AI and its carbon footprint due to the increasing energy demand of AI models.
• Model Selection: Hyperparameter optimization and neural architecture search are energy intensive processes. Methods for obtaining inherently efficient architectures and model configurations will be discussed (AutoML).
• Low-resource model training: Training of AI models incurs recurrent costs due to iterative optimization over large datasets. During the course, we will show several techniques to make model training more efficient.
• Efficient AI deployment: Trained AI models are often resource intensive to run. When deployed on edge devices (mobile phones) they have additional resource constraints. Converting AI models into efficient instances for deployment will be explored (TinyML).
• Documentation: Presentation of best practices in measuring and reporting carbon footprint and related metrics to facilitate transparency, compliance and measurable improvement.
• Climate-friendly AI applications: Applications of AI aligned with climate change mitigation, as well as environmental conservation and preservation, will be showcased.

• Carbon footprint and energy consumption estimation tools, including Carbontracker.
• Hyperparameter optimization techniques for AutoML.
• Multi-objective optimization methods to achieve trade-off between performance and energy efficiency.
• Small and efficient trained models for Natural Language Processing and Computer Vision, as well as methods to create such models from larger ones.
• Model cards for standardized documentation of carbon footprint and energy consumption.

We will primarily be working with Python using PyTorch as the deep learning framework; however, all techniques that are covered are easily applicable to other frameworks.

The course is strictly focused on Machine Learning and Big Data Analysis, so a prerequisite is that you have a background in developing and/or deploying Machine Learning models. This course assumes you have studied to at least Bachelor degree level and/or have several years of AI/ML experience.