WEL-T starting grant "Embodied Excitability" (2023-2027)

“Intelligent” machines are irreplaceable pieces of our society. But the digital Von Neumann architecture lying at their core makes them extremely energy hungry. They are also rigidly optimized and fragile to perturbations, which makes them hardly reusable. If we want our machines to be more sustainable, we need to rethink the way in which we design them. Biology can be a great source of inspiration by showing us numerous examples of intelligence based on adaptability and energy efficiency, instead of performance. 

Biological intelligence is analog, embodied, and dynamical. A governing principle of biological intelligence is excitability. 

Neuromorphic engineering aims at using dynamics and excitability to design analog chips that depart from the Von Neumann architecture through “in-memory computation”, which might be key for increased adaptability and energy efficiency. But neuromorphic engineering is still far from reaching its full potential. We suggest that this is due to the lack of a principled theoretical framework for neuromorphic engineering that provides understanding of how embodied intelligence based on excitability works and how we can reproduce it in machines.

In this project, we introduce the principle of Embodied Excitability as the synthesis of the defining properties of biological intelligence and use it to develop a theoretical framework for the analysis and design of neuromorphic adaptive agents. Under this principle, we design adaptive neuromorphic sensors and actuators with in-sensor decision-making functions, as those underlying biological attention, and couple them in embodied sensorimotor loops controlling autonomous agents.

Our results might lead to energy-efficient intelligent sensors and actuators that in the future might be relevant for sustainable developments in the unmanned transport sector.

Partners:

• University of Zurich & Institute of Neuroinformatics

• Groningen University

• Italian Institute of Technology