Energy-Efficient AI: Why brains use much less power than AI, and how this might help us improve current AI
Brains are incredibly energy-efficient, with each connection using thousands to millions of times less energy than a connection weight a modern Large Language Model running on GPUs. This talk looks at why this is the case, focusing on how brains differ from traditional digital hardware. Furthermore, we look at how modern understanding of how brains represent Time and Space (Grid Cells and Time Cells) has led to new AI algorithms that improve efficiency. However, to fully take advantage of these improvements, we also need to rethink how we design computer hardware, leading to more "neuromorphic" approaches.
About the Speaker
Terry Stewart
Associate Research Officer, Centre for Theoretical Neuroscience, University of Waterloo
Terry Stewart is currently an NRC Canada Associate Research Officer and adjunct professor with the Dept of Psychology at the University of Waterloo. He is a former member of the Computational Neuroscience Research group at the University of Waterloo working with Chris Eliasmith on the development of Spaun and the Nengo software.