I just now posted a small research / ML interp study on symmetries inside the neural nets for an open source Go-playing program that I maintain ("KataGo"). The rules of Go are completely symmetric under rotation/reflection, but such symmetry is not enforced in the models - the only thing we do for that is stochastic 8-fold data augmentation during training, randomizing the spatial orientation of each batch. To what degree do superhuman-strength Go-playing neural nets automatically l...