Exploring Symmetries in a Go Network's Neural Architecture
A study on neural symmetries in KataGo reveals insights into Go-playing AI's internal representations.
A recent study investigates the symmetries within the neural networks of KataGo, an open-source Go-playing program. While the rules of Go are symmetric, this symmetry is not enforced in the models. The research examines how effectively superhuman Go-playing neural networks learn to represent the board independently of its orientation, revealing interesting insights into their internal processes. This study contributes to the broader field of interpretability in machine learning.