Neuromorphic Intelligence: Bio-Inspired Algorithms for Hardware-Native Neural Computing

Abstract: The brain is efficient, and modern AI is not. This mismatch is a matter of design. The brain utilizes analog computations distributed across sophisticated organic network topologies, whereas modern AI implements the backpropagation algorithm on digital GPUs supported by Von Neumann architectures. There are a growing number of emerging hardware that implement brain-like networks and operations, but there is not yet a clear neuromorphic algorithm to match the backpropagation-level performance required for impactful intelligence. This dissertation explores a number of novel and known candidates to support intelligence in neuromorphic hardware, and culminates in a bio-inspired learning algorithm that consistently performs within 1% of backpropagation using only hardware-friendly operations. This result may inform advancements in machine learning, specialized hardware design, and even neuroscience.

Bio: Ryan produced this work in collaboration with the National Institute of Standards and Technology (NIST) and the University of Colorado Boulder (CU Boulder). He is advised by Bradley Hayes at CU, and Adam McCaughan and Sonia Buckley at NIST. Ryan has encountered a diverse academic background on his way to this defense of a PhD in Computer Science. His undergraduate studies were in Astrophysics and Philosophy, and he holds a graduate degree in Artificial Intelligence. The common theme is an interest in exploring fundamental questions, which has most recently coalesced into the targeted pursuit of how natural intelligence might be realized in physical systems. This same explorative streak extends to Ryan's personal life, where he enjoys adventures of all sorts with his life-partner Anna, and especially long-term backpacking.