Research map / Label-efficient and robust learning / Recognition and long-tail
On the Intrinsic Dimensionality of Image Representations
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Map SphereFace to 10 dimensions with DeepMDS instead of retaining its 512-dimensional representation | SphereFace · LFW (BLUFR) | TAR @ 0.1% FAR: 96.74% -> 92.33% | Hurt |
| Use DeepMDS instead of PCA, Isomap, or DAE for a 10-dimensional mapping | SphereFace · LFW (BLUFR) | TAR @ 0.1% FAR: 16.04% -> 92.33%; 77.31% -> 92.33%; 6.73% -> 92.33% | Helped |
| Compress ResNet-34 features from 512 to 19 dimensions with DeepMDS | ResNet-34 with a linear classifier · ImageNet-100 | Top-1 accuracy: 80.0 -> 68.0 | Hurt |
| Use DeepMDS rather than DAE for 19-dimensional features | ResNet-34 with a linear classifier · ImageNet-100 | Top-1 accuracy: 50.2 -> 68.0 | Helped |
| Compress ResNet-34 features from 512 to 19 dimensions with DeepMDS | ResNet-34 with a k-NN classifier · ImageNet-100 | Top-1 accuracy: 83.4 -> 77.0 | Hurt |
| Use DeepMDS rather than DAE for 19-dimensional features | ResNet-34 with a k-NN classifier · ImageNet-100 | Top-1 accuracy: 73.4 -> 77.0 | Helped |
| Compress SphereFace from 512 to 12 dimensions with DeepMDS | SphereFace · IJB-C | TAR @ 0.1% FAR: 69.32% -> 62.25% | Hurt |
| Train the ambient-to-intrinsic mapping directly instead of stagewise (ablation) | DeepMDS · LFW (BLUFR) | TAR @ 0.1% FAR: 92.33 -> 80.25 | Hurt |
| Add intermediate supervision to direct mapping (ablation) | DeepMDS · LFW (BLUFR) | TAR @ 0.1% FAR: 80.25 -> 86.15 | Helped |
| Use stagewise training instead of direct mapping with intermediate supervision (ablation) | DeepMDS · LFW (BLUFR) | TAR @ 0.1% FAR: 86.15 -> 92.33 | Helped |
| Add fine-tuning after stagewise training (ablation) | DeepMDS · LFW (BLUFR) | TAR @ 0.1% FAR: 92.33 -> 90.42 | Hurt |
Closest papers by meaning
- Intrinsic Dimension Correlation: uncovering nonlinear connections in multimodal representations (ICLR 2025 · similarity 0.86)
- The Effect of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images (ICLR 2024 · similarity 0.85)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.83)
- Exploring Intrinsic Dimension for Vision-Language Model Pruning (ICML 2024 · similarity 0.83)
- A Wiener Process Perspective on Local Intrinsic Dimension Estimation Methods (AAAI 2025 · similarity 0.82)
- A Geometric View of Data Complexity: Efficient Local Intrinsic Dimension Estimation with Diffusion Models (NeurIPS 2024 · similarity 0.82)
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.