Sustainable ML
Machine learning (ML) is a critical component of artificial intelligence (AI) development, allowing for complex tasks to be accomplished by learning from data. However, the energy and resource consumption of digital infrastructure, such as data centers and networks, may lead to adverse environmental effects. About 5-9% of the world’s total electricity production is used by information and communication technology, which may rise to 20% in 2030. For example, the development of large language models has been associated with a doubling of the computing requirements every two months. It is important to balance the benefits of ML with its sustainability challenges and prioritize environmentally and socially responsible AI development. More resource-efficient co-design and use of heterogeneous hardware and algorithms for robust and explainable distributed AI in the cloud-to-edge continuum, as well as better software engineering processes are required to make the development of AI sustainable and widely accessible for the long-term welfare of society. Our work focuses on developing programming, interoperability and distributed machine learning concepts to enable a next generation of energy-efficient, explainable and robust AI based on neuromorphic, quantum and digital hardware forming a sustainable cloud-to-edge continuum.
We aim to balance the potential of Machine Learning with sustainability challenges in the development of AI through resource-efficient solutions.
Projects
MSc theses
2017 Andreas Ternstedt, Pattern recognition with spiking neural networks and the ROLLS low-power online learning neuromorphic processor, link to thesis External link., news article External link.
2018 Mattias Nilsson, Monte Carlo Optimization of Neuromorphic Cricket Auditory Feature Detection Circuits in the Dynap-SE Processor, link to thesis External link.
2019 Oskar Öberg, Critical Branching Regulation of the E-I Net Spiking Neural Network Model, link to thesis External link., collaboration External link.
2021 Olof Johansson, Training of Object Detection Spiking Neural Networks for Event-Based Vision, link to thesis External link.
2022 Kim Petersson Steenari, A neuromorphic approach for edge use allocation, link to thesis External link., collaboration External link.
PhD theses
2023 Mattias Nilsson Event-Driven Architectures for Heterogeneous Neuromorphic Computing Systems link External link.
Machine learning (ML) plays a central role in the development of artificial intelligence (AI) as ML models enable AI systems to perform complex tasks by learning from data, such as interpreting language and generating speech to answer questions. AI-generated information and automation solutions can help us to develop more sustainable processes and behaviour but also contributes considerably to the electric energy and resource consumption of digital infrastructure such as data centres and networks. This creates risks of adverse environmental effects. It has been estimated that 5-9% of the world’s total electricity production is used by information and communication technology, which may rise to 20% in 2030. It is important to balance the potential benefits of ML with its sustainability challenges, and to prioritize the development of AI that is environmentally and socially responsible. More resource-efficient ML hardware, algorithms and software engineering concepts are required to make the development of AI sustainable and widely accessible for the welfare of society.
Updated: