Public defence: Tom Eivind Harlem Glover
Tom Eivind Harlem Glover will defend his thesis “Harvesting the computation of Elementary Cellular Automata with Reservoir Computing” for the PhD in Engineering Science.
Tom Eivind Harlem Glover will defend his thesis “Harvesting the computation of Elementary Cellular Automata with Reservoir Computing” for the PhD in Engineering Science.
The dissertation is available in NVA Digital Archive.
The trial lecture and public defence will also be streamed live.
Ordinary opponents:
Leader of the public defence: Professor and Head of the Department of Computer Science André Brodtkorb, OsloMet.
Main supervisor: Professor Stefano Nichele, Østfold University Collage and OsloMet.
Co-supervisors:
Modern Artificial Intelligence (AI) systems achieve impressive results, but they consume a significant amount of energy. In some cases, smarter AI can reduce overall energy use, but it can also drive up consumption by making it easier to deploy at scale. This creates a complex relationship between performance and sustainability. In addition, the high energy cost limits the capabilities of the AI, it makes them difficult to run in small, battery-powered devices like smartwatches or in other settings where energy must be conserved. For both reasons, a more energy-efficient AI alternative could be helpful.
This research explores an alternative approach that could lead to more energy-efficient AI by drawing inspiration from biology. The method is called Reservoir Computing with Cellular Automata (ReCA). It combines two components: reservoir computing, which is inspired by one of the many ways the brain might process information efficiently, and cellular automata, a simple model inspired by biological systems such as biological development and self-replication. The result is a system that is promising for AI tasks where energy use needs to be minimised, such as Edge AI. This could also enable valuable features, such as improved privacy, since personal data can be processed locally on the device instead of being sent to the cloud.
This project examined how different configurations and settings affect ReCA’s behaviour. We found that small changes in the setup can result in significant differences in performance. Among the alternatives tested, elementary cellular automata showed particular promise. In particular, ReCA performs well on tasks where information is structured locally, such as image recognition. However, it has not yet performed well on tasks that require a more global perspective of the data.
Although ReCA is still in an early stage, the results show that it has clear potential as a tool for building energy-efficient.