pipette
ENEnglish

Asymmetry Tunneling Beyond Adaptive Equivariance Relaxation

Eunho Jeong, Joon Jang, Hyeonjin Kim

Preprint

En palabras de los autores

Equivariant neural networks encode prescribed symmetries as powerful inductive biases, yet real-world data may exhibit symmetry-breaking structure that these priors cannot fully capture. Prior work has largely addressed this challenge through model-side equivariance relaxation, seeking an appropriate level of equivariance for the data. This raises a complementary question: must equivariance relaxation be adaptively controlled? To investigate this question, we introduce a residual path alongside a strictly equivariant model and deliberately restrict their interface to examine the transmission of task-relevant structure. Across five benchmarks, we observe both residual collapse and survival despite these restrictions; we call the latter phenomenon asymmetry tunneling. We further characterize this phenomenon through four complementary diagnostics. Our results suggest that adaptive control of equivariance relaxation is not a prerequisite for capturing potential symmetry-breaking structure.

Resultado principalEl resumen no menciona limitaciones.

Apareció: jueves, 22 de octubre. arXiv. Preprint, todavía sin revisión por pares.