Tag
A study on how KataGo's Go-playing neural networks internally represent board symmetries, finding that the model learns partially symmetric concepts despite only stochastic 8-fold data augmentation, with one unexpected result.