Research map / Learning theory and algorithms / Matrix and tensor methods
Local Hessian Spectral Filtering for Robust Intrinsic Dimension Estimation
GENKI OSADA
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Hessian spectral filtering (LHSD m=5) instead of FLIPD | FLIPD · Synthetic manifolds average (9 settings) | MAE: 307.4 -> 6.6 | Improved |
| use LHSD instead of FLIPD in high co-dimension | FLIPD · F^10+80+200 in R^3072 | MAE: 1240.89 -> 5.19 | Improved |
| use LHSD instead of FLIPD on nonlinear d=900 manifold | FLIPD · F^900 in R^3072 | MAE: 782.5 -> 18.79 | Improved |
| increase Lanczos steps from m=2 to m=5 (ablation) | LHSD · Synthetic manifolds average | MAE: 20.7 -> 6.6 | Improved |
| increase Lanczos steps from m=5 to 10 or 20 (ablation) | LHSD · Synthetic manifolds | MAE | No clear effect |
| increase Lanczos steps from m=2 to m=5 (ablation) | LHSD · IDR benchmarks runtime | runtime (ms): 480 -> 1,200 | Hurt |
| use Hessian spectral filtering (LHSD m=2) instead of FLIPD | FLIPD · IDR benchmark runtime | runtime (ms): 120,000 -> 480 | Improved |
| apply Hutchinson estimation (K=8) to FLIPD (ablation) | FLIPD · IDR benchmark | MAE: error > 12.9 | Hurt |
| use LHSD instead of FLIPD under IDR embedding D=784 | FLIPD · IDR Funnel | MAE: 1.606 -> 0.661 | Improved |
| train score model on 500k instead of 50k samples (ablation) | LHSD · L^10+80+200 D=1024 | MAE: 4.16 -> 3.47 | Improved |
| train score model on 500k instead of 50k samples (ablation) | LHSD · F^10+80+200 D=1024 | MAE: 7.52 -> 6.90 | Improved |
| vary filter steepness p (ablation) | LHSD · L^900 in R^3072 | MAE | No clear effect |
Closest papers by meaning
- A Wiener Process Perspective on Local Intrinsic Dimension Estimation Methods (AAAI 2025 · similarity 0.88)
- A Geometric View of Data Complexity: Efficient Local Intrinsic Dimension Estimation with Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Diffusion Models Encode the Intrinsic Dimension of Data Manifolds (ICML 2024 · similarity 0.87)
- The Noisy Laplacian: a Threshold Phenomenon for Non-Linear Dimension Reduction (ICML 2025 · similarity 0.86)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.85)
- On gauge freedom, conservativity and intrinsic dimensionality estimation in diffusion models (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
