'''群运算(Group Operations):'''这是模运算的推广。Chughtai 等人<ref>{{cite journal | author1=Chughtai B | author2=Chan L | author3=Nanda N | year=2023 | title=A toy model of universality: Reverse engineering how networks learn group operations | journal=International Conference on Machine Learning (ICML), PMLR | volume= | issue= | page= | url= }}</ref>训练MLP模型在阶数为<math>∣G∣=n</math>的有限群上执行群组合运算,发现模型表现出"顿悟"现象。模型输入为有序对 <math>(a,b)</math>(其中<math>a, b \in G</math>),训练目标是预测群元素 <math>c=ab</math>。值得注意的是,Nanda等人<ref name=":0" />提出的任务是该任务的特例——因为模113加法运算等价于113阶循环群的组合运算。 | '''群运算(Group Operations):'''这是模运算的推广。Chughtai 等人<ref>{{cite journal | author1=Chughtai B | author2=Chan L | author3=Nanda N | year=2023 | title=A toy model of universality: Reverse engineering how networks learn group operations | journal=International Conference on Machine Learning (ICML), PMLR | volume= | issue= | page= | url= }}</ref>训练MLP模型在阶数为<math>∣G∣=n</math>的有限群上执行群组合运算,发现模型表现出"顿悟"现象。模型输入为有序对 <math>(a,b)</math>(其中<math>a, b \in G</math>),训练目标是预测群元素 <math>c=ab</math>。值得注意的是,Nanda等人<ref name=":0" />提出的任务是该任务的特例——因为模113加法运算等价于113阶循环群的组合运算。 |