Machine learning and NLP systems have become huge and demand vast amounts of resources to train and to predict. But we can do very well with much less than some large language models require. We investigate measuring and reducing the amount of data, energy, and time needed to do machine learning and language processing.

• Hyperparameter Power Impact in Transformer Language Model Training. 2021.
• Optimal Size-Performance Tradeoffs: Weighing PoS Tagger Models. 2021.
• Power Consumption Variation over Activation Functions. 2020.

• Climate change and large-scale language models: a panel discussion, Swedish Research Institute NLP Webinars, 2021
• Green Natural Language Processing, Swedish Language Technology Conference opening keynote, 2020
• Green AI, Data Science & Society invited talk, ITU, 2020

Flick through this brief summary presentation on Green NLP:

• This Powerful AI Technique Led to Clashes at Google and Fierce Debate in Tech. Here’s Why. - Morning Brew, 2021.03.21

• Featherweight NLP: Tolerable Lower Limits for Size of Transformer Language Models. Magnus Malthe Jacobsen, Mikkel Hooge Sørensen, 2021, MSc
• Low-carbon NLP Models. Lucas Høyberg Puvis de Chavannes, Mads Guldborg Kjeldgaard Kongsbak, Timmie Mikkel Rantzau Lagermann, 2021, MSc
• Smart Charging of Electric Scooters Using Real-Time Electricity GHG Emissions. Markus Killendahl, Felix Dawids Qvist, 2021, BSc