Low-dimensional topology of deep neural networks
深度神经网络的低维拓扑
机构 * University of Chicago(芝加哥大学)
AI总结 通过限制每层宽度为3,研究前馈网络、ResNet和Transformer如何改变低维拓扑不变量(如环绕数),揭示不同架构的表达能力差异。
Comments 24 pages, 11 figures; published at ICML 2026 (PMLR 306). Wording polished for clarity; technical content and results unchanged