Hyperbolic Manifold Domain Map

核心定义

Hyperbolic Geometry & Manifold Learning = 将深度学习扩展到 curved spaces（双曲空间、Riemannian manifold），利用非欧几何的 exponential volume growth 特性处理 hierarchical data、scale-free networks、constrained optimization。

技术架构

mindmap
  root((Hyperbolic Manifold))
    Geometry
      Poincaré Ball
      Lorentz Model
      Riemannian Manifold
    Application
      Hierarchical Embedding
      Scale-free GNN
      Constrained Optimization
    Challenge
      Optimization Cost
      Architecture Extension
      Scalability

研究路线

1. Hyperbolic Embeddings (Foundational)

里程碑: 1700-PoincareEmbeddings (🔥 Rating 5)

核心洞察: Hyperbolic volume exponential ≈ tree exponential

关键结果: 5-dim ≈ 100-dim，20x+ dimensionality reduction

应用: Taxonomy embedding, knowledge graphs

2. Hyperbolic Neural Networks

Survey: 2400-HyperbolicNeuralNetworksSurvey (Rating 4)

核心思路: Neural layers in hyperbolic space

技术挑战:

Linear ops 需要 tangent space 映射
Attention extension 不成熟
初始化敏感

应用: KG embeddings, word embeddings, hierarchical classification

3. Hyperbolic Graph Neural Networks

代表: 2400-HyperbolicGNN (Rating 3)

核心思路: Scale-free networks benefit from hyperbolic geometry

设计: Hyperbolic message passing + attention

优势: Hub importance captured, 参数效率高

局限: Tangent space mapping cost 高

4. Riemannian Optimization

代表: 2400-RiemannianOptimization (Rating 3)

核心思路: Gradient descent on manifold constraints

Manifolds: Stiefel (orthogonal), Grassmann (low-rank), Positive Definite

应用: Orthogonal RNNs, low-rank training, covariance estimation

近期进展:

2502-AcceleratedRiemannianOptimization (Rating 3) - 多 constraints 同时优化
2502-RiemannianFixedRank (Rating 3) - Fixed-rank manifold metrics 分析

5. Lorentz Graph Neural Networks

代表: 2405-LorentzGNN (Rating 3)

核心思路: Lorentz-equivariant message passing for physics

亮点: Equivariance 提升泛化，减少 data 需求

6. Hyperbolic Contrastive Learning

代表: 2501-HyperbolicGraphContrastive (Rating 3)

核心思路: Hierarchical positive sampling in hyperbolic space

应用: Node classification SOTA, graph property prediction

7. Scalable Hyperbolic Knowledge Graphs

代表: 2502-ScalableHyperbolicKG (Rating 3)

核心思路: Mini-batch training for 百万实体 KG

工程突破: Scalability bottleneck 解决

8. Hyperbolic Attention Networks

代表: 2503-HyperbolicAttention (Rating 3)

核心思路: Distance-based attention for long-range dependencies

应用: Document translation, long-context LM

Benchmarks

Benchmark	类型	SOTA
WordNet	Taxonomy	Poincaré Embeddings
Freebase	KG	Hyperbolic KG
Citation Networks	Scale-free	Hyperbolic GNN

关键洞察

Pattern 1: Hyperbolic ≈ Hierarchy

Exponential volume growth 天然匹配 hierarchical/tree structures

Pattern 2: 维度效率惊人

Poincaré 证明 5-dim hyperbolic ≈ 100-dim Euclidean

Pattern 3: Scale-free Networks Benefit

Power-law degree graphs 在 hyperbolic space 表现优异

Pattern 4: Optimization 是瓶颈

Riemannian operations cost 高，限制大规模应用

Pattern 5: Curvature Learning 是趋势

Data-adaptive manifold curvature estimation 正在兴起

待解决问题

大规模 Scalability（distributed Riemannian optimization）
Attention in Hyperbolic Space（Transformer extension）
Curvature Learning（自动选择 curvature）
Dynamic Hierarchies 处理
Multimodal Hyperbolic（VLM/Agent 应用）
Optimization Speed 加速

可视化演示

🌐 在线浏览 HTML 演示 — 杂志风格翻页展示

下一步

方向	Action
Foundational	精读 Poincaré Embeddings 论文了解 RSGD 实现
GNN	研究 Hyperbolic GNN 的 attention mechanism
Optimization	研究 Natural Gradient 与 Information Geometry 关联
Application	探索 VLM/Agent 的 hyperbolic embedding 潜力

DeepRead

Explorer