Independent Research Lab
Coming
soon.
We are setting up a small lab for research on machine learning that respects structure — symmetry, geometry, and physics. 我们正在搭建一个小型研究组，做「尊重结构」的机器学习：对称性、几何与物理。
Research plan 研究计划 · 2026–2027
Geometric Deep Learning面向几何的深度学习
Symmetry as an inductive bias. We study equivariant networks on graphs, meshes, point clouds and manifolds — building models whose behaviour is dictated by the transformation group of the domain rather than by memorised patterns.
Structure-Preserving Learning结构保持学习
Learning systems that do not discard the physics. Differentiable simulation, physics-informed objectives, and discrete differential operators on simplicial complexes — with conservation properties verified, not assumed.
Efficient Computation on Irregular Data不规则数据上的高效计算
Most real geometry is irregular. We work on sparse operators, subquadratic message passing, and hierarchical coarsening, so that equivariant models stay usable at industrial mesh scale.
Theory & Evaluation理论与评测
Claims about symmetry need instruments. We maintain honest benchmarks: expressivity and over-squashing, symmetry-breaking diagnostics, and reproducible geometric baselines with published failure cases.
Method
Small, honest releases. Technical notes before papers; code before claims.
Output
Open notes, reference implementations, and reproducible geometric benchmarks.
Timeline
Q4 2026 foundations → 2027 Q1 first technical notes → 2027 Q2 reference code.
Get in touch 联系
If you work on geometric learning, equivariant architectures, or physics-informed models — we would like to hear from you. 如果你在做几何学习、等变架构或物理信息模型，欢迎聊聊。