Neural Networks Provably Learn Spectral Representations for Group Composition

Arxiv CS.LG (machine learning)

The paper proves that two-layer neural networks can learn spectral representations for group composition tasks, providing theoretical insight into structured internal representations. The result holds for finite samples and shows that the learned representations align with the group's irreducible representations.

Arxiv CS.LG (machine learning)·Jianliang HeLeda WangFengzhuo ZhangSiyu ChenZhuoran Yang

8 days ago