Decarbonising AI with event-based neural networks

The artificial neural networks underlying modern AI are very computing-intensive and hence take a lot of energy to run. In contrast, the human brain has a power budget of only about 20 Watts.

Some of the brain's efficiency is due to sparse activation and communication through spikes. This is the motivation for using spiking neural networks in novel energy-efficient neuromorphic hardware.

In this lecture, I will first give a brief overview of today's AI systems and acute issues that we are facing with them. I will then discuss how event-based processing in spiking neural networks may offer a path towards addressing the issue of modern AI's excessive energy consumption. I will briefly discuss some theory and then show results of our recent work that show competitive performance on keyword recognition tasks, beneficial scaling with the length of input sequences, and 2000 times energy savings when deploying trained networks on the Intel Loihi 2 neuromorphic system. I will conclude with some reflections on interpreting interactions with AI systems.