What AI ‘Sees’ and Why It Matters

Hyperbolic space vs. Euclidean space → Hyperbolic space stabilizes hierarchical structures better, improving image classification accuracy. Probabilistic vs. deterministic hierarchy trees → Probabilistic modeling outperforms deterministic methods, enhancing recognition through Mixture-of-Gaussians (MoG). Hierarchy width (N) & depth (L) → Best results on ImageNet-1K achieved at N=32, L=4, balancing fine-grained representation and optimization.