Abstract: The explosively growing computational needs of modern AI models combined with the slowdown of Moore’s Law creates a multitude of issues, from a lack of reproducibility and open access to large carbon footprints. One path to alleviate these issues seeks to improve computational efficiency for AI by creating systems which utilize biological principles of computation discovered in the brain. ​‘Phasor networks’ are one such system, in which activations in a neural network represent the relative phase of oscillators.

This talk will present an overview of this interpretation of neural networks and the advantages which it can bring, including the ability to manipulate neural layers algebraically and links to temporal, analog computation.