SOMA

The talk surveys alternatives to GPU based computing, arguing that hardware and learning algorithms need to be designed together. The host presents zero order optimization and sharded model assemblies as ways to train without backpropagation, then introduces optical computing and neuromorphic hardware as possible energy efficient substrates. The optical computing presentation explains how light can perform parallel linear operations and describes a fixed optical system for diffusion image generation, along with its limits in memory, nonlinear operations, and data conversion. The neuromorphic presentations cover brain inspired circuit design and a system that trains cultured brain cells to play Doom through learned stimulation, decoding, and feedback.