Embedding boolean circuits in a neural network
Summary
Edward Benson explores embedding a boolean circuit inside a neural network, detailing how to represent boolean operators as neurons and connect a boolean sidecar to a trained network. The post introduces the concepts of probes (ingress) and stims (egress) to translate between real-valued neural signals and boolean signals, aiming to create a hybrid architecture with safety and interpretability considerations. This is presented as a thought experiment and planning groundwork for a more involved project.