Co-Designing Hardware and Algorithms: Lessons from the Brain

Abstract: Neuromorphic computing aims to emulate the brain’s computational architecture with low energy consumption, offering a way to enhance the performance of modern, power-hungry artificial intelligence (AI) systems. Achieving brain-like cognition and maximizing neuromorphic benefits requires complex neurons, dense connectivity, heterogeneous compute components, scalability with trillions of learnable parameters and novel algorithms.

This necessitates a co-design approach spanning algorithms, architectures, systems, circuits and devices, along with heterogeneous integration techniques. There is a significant opportunity to design efficient next-generation AI that leverages underlying hardware much like biological co-design seen in the brain.

In this talk, I will present our recent work on developing a hardware generalizable method for training physical neural networks, which utilizes nonlinearities and nonidealities to perform efficient computation.

Refreshments will be served.

Series: The Argonne Microelectronics Institute colloquium series invites leaders in the field from academia, national labs and industry to present their forefront research and perspectives. Join us in person or online to learn and connect.

If you have questions about the colloquium series or would like to suggest or host a future speaker, contact Dan Durham (durhamd@​anl.​gov).