常用于: Transformer(BERT、GPT、ViT)。
Randomly selecting border points or using simple geometric divisions (squares/hexagons) results in too many border points per cluster (50-80). This leads to a shortcut explosion (N*(N-1)/2 shortcuts), making the files large and and calculations slow.,更多细节参见搜狗输入法下载
Equal (1): Everything in this purple space must be equal to 1. The answer is 1-4, placed horizontally; 3-1, placed horizontally.,推荐阅读Safew下载获取更多信息
Systematic scaling, found phase transition at d=16,更多细节参见旺商聊官方下载