New research project in Deep Learning very beneficial for the environment
A new research project, which is funded by the Swedish Research Council, will develop energy-efficient Deep Learning models to help reduce the carbon footprint. Today, Deep Learning (DL) is widely used in learning and interpreting large amounts of data, such as in health monitoring and robotics, but also in social media networking. Today's DL models require a high energy consumption, which leads to substantial environmental costs. For instance, a DL model can generate carbon dioxide emissions equivalent to the total lifetime carbon dioxide emissions of five cars.
“It appears that Deep Learning will account for a substantial part of carbon dioxide production during the next decade. We wish to change that,” says Professor Masoud Daneshtalab, who conducts research in deep learning and AI acceleration at Mälardalen University.
Deep Learning is a type of machine learning and artificial intelligence that imitates how the human brain works and learns. It is an important element of data science and includes statistics and predictive modelling. Deep learning is useful, for example, in automated driving to detect pedestrians and reduce accidents.
“In AI research, researchers have focused mainly on improving the accuracy and reliability of the results and have paid relatively little attention to energy efficiency. In this research project, we can contribute to the UN's sustainable development goals while not jeopardising the continued development of deep learning,” says Masoud Daneshtalab.
Scalable energy-efficient DL models
The research project is being conducted under the name GreenDL: Green deep learning for edge devices and will run for four years from 2022 to 2025. The aim of the project is, among others, to develop new theoretical foundations and algorithms to design scalable and energy-efficient DL models.
To create the greatest impact, knowledge about the importance of using green DL models will be communicated in society.
“We will make the source code available in the software from the beginning by creating an open-source project where industry and other researchers can access our results,” says Masoud Daneshtalab.
Masoud Daneshtalab previously worked at the University of Turku, Finland and the Royal Institute of Technology (KTH) where he discovered artificial intelligence and machine learning when he wrote his degree project on a brain-like computer paradigm in 2013. He spent two years at KTH developing a customisable hardware implementation of a scalable brain-inspired computational architecture. He has been working at Mälardalen University since 2016 and is a member of the internationally renowned Embedded Systems research specialisation where he co-leads the Heterogeneous System research group, hardware/software co-design.
What possibilities do you see for AI in the future?
“My vision is to create a customised hardware design with the emerging memory technology to develop a complete real-time capable computing device that matches the human brain in terms of the numbers of units and connections, in the size of a Smart watch.”
Masoud Daneshtalab feels at home at Mälardalen University.
“I really enjoy both researching and teaching at the University. But the freedom and support I get to conduct my research is what I love most about working here.”