Research map / Graph learning / Graph algorithms and clustering
Theoretical Analysis of the Optimal Free Responses of Graph-Based SFA for the Design of Training Graphs
Alberto N. Escalante-B., Laurenz Wiskott
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 |
|---|---|---|---|
| add 9 auxiliary cosine labels to the gender label (10 target labels) (ablation) | ELL^g graph, HGSFA · Gender estimation, artificial faces (soft GC, 1F) | RMSE: 0.364 -> 0.353 | Improved |
| add 39 auxiliary labels (40 target labels) (ablation) | ELL^g graph, HGSFA · Gender estimation, artificial faces (soft GC, 1F) | RMSE: 0.364 -> 0.346 | Improved |
| raise target labels from 40 to 50 (ablation) | ELL^g graph, HGSFA · Gender estimation, artificial faces (soft GC, 1F) | RMSE: 0.346 -> 0.351 | Hurt |
| replace linear scaling mapping by soft Gaussian classifier mapping (ablation) | ELL^g-40 · Gender estimation, artificial faces | RMSE: 0.368 -> 0.346 | Improved |
| use ELL^g-40 instead of serial graph | GSFA training graph · Gender estimation, artificial faces (soft GC, 1F) | RMSE: 0.351 -> 0.346 | Improved |
| code gender and color jointly (2x5 labels) instead of gender only (ablation) | ELL graph · Gender estimation, artificial faces (soft GC, 3F) | RMSE: 0.345 -> 0.277 | Improved |
| increase auxiliary labels in joint gender-color graph with 1 feature (ablation) | ELL^{g,c} · Gender estimation, artificial faces (soft GC, 1F) | RMSE: 0.289 -> 0.499 | Hurt |
| code gender and color jointly instead of color only (3 features) (ablation) | ELL graph · Color estimation, artificial faces (soft GC, 3F) | RMSE: 1.867 -> 1.140 | Improved |
| add 30 auxiliary labels (compact+31 vs compact+5) at d=5 (ablation) | compact graph · GTSRB 32 classes, d=5 | error rate (%): 11.67% -> 12.42% | Hurt |
| use compact+5 graph instead of clustered graph at d=5 | GSFA · GTSRB 32 classes, d=5 | error rate (%): 29.74% -> 11.67% | Improved |
| use compact+31 graph instead of clustered graph at d=31 | GSFA · GTSRB 32 classes, d=31 | error rate (%): 2.89% -> 2.89% | No clear effect |
| preserve more than 5 features with compact+5 graph (ablation) | compact+5 · GTSRB 32 classes | error rate (%) | Hurt |
Closest papers by meaning
- Labeled Graph Kernel for Behavior Analysis (TPAMI 2016 · similarity 0.84)
- Source-Free Progressive Graph Learning for Open-Set Domain Adaptation (TPAMI 2023 · similarity 0.82)
- Graphical Lasso and Thresholding: Equivalence and Closed-form Solutions (JMLR 2019 · similarity 0.82)
- Generalization Analysis for Label-Specific Representation Learning (NeurIPS 2024 · similarity 0.81)
- GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers (ICML 2025 · similarity 0.81)
- Designing Specialized Two-Dimensional Graph Spectral Filters for Spatial-Temporal Graph Modeling (AAAI 2025 · similarity 0.81)
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.
