Research map / Label-efficient and robust learning / Self-supervised and contrastive
Inverse Problems Leveraging Pre-trained Contrastive Representations
Read the paperCheck 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 |
|---|---|---|---|
| Train a contrastive student to match clean CLIP representations from corrupted inputs, then use a linear probe | ImageNet-100 classification with random masking · ImageNet-100 val, Random Mask 90% | top-1 accuracy: 74.12$\pm$0.09 -> 82.96$\pm$0.08 | Helped |
| Train a contrastive student to match clean CLIP representations from corrupted inputs, then use a linear probe | ImageNet-100 classification with additive Gaussian noise · ImageNet-100 val, Gaussian Noise σ = 0.5 | top-1 accuracy: 71.43$\pm$0.14 -> 76.23$\pm$0.10 | Helped |
| Train a contrastive student to match clean CLIP representations from corrupted inputs, then use a linear probe | ImageNet-100 classification with Gaussian blur · ImageNet-100 val, Gaussian Blur n = 37 | top-1 accuracy: 68.94 -> 77.80 | Helped |
| Use only 10% of labeled data for the linear probe instead of all labeled data | Contrastively trained CLIP robust encoder · ImageNet-100 val, Random Mask 90% | top-1 accuracy: 82.96$\pm$0.08 -> 79.87$\pm$0.12 | Hurt |
| Use only 10% of labeled data for the linear probe instead of all labeled data | Contrastively trained CLIP robust encoder · ImageNet-100 val, Random Mask 90% | top-1 accuracy versus supervised baseline: 74.12$\pm$0.09 -> 79.87$\pm$0.12 | Helped |
| Use a contrastive uniformity term rather than MSE alone (ablation) | Robust encoder trained on ImageNet-100 random masking · ImageNet-100 val, masking levels 90–99% | top-1 accuracy | Helped |
| Replace student-versus-teacher comparisons with student-versus-student comparisons in the uniformity term (ablation) | Contrastive robust encoder · ImageNet-100 val, masking levels 90–99% | top-1 accuracy | No clear effect |
| Use NT-XEnt comparisons across student and teacher representations instead of the paper's student-versus-teacher uniformity term (ablation) | Contrastive robust encoder · ImageNet-100 val, masking levels 90–99% | top-1 accuracy | Hurt |
| Transfer the robust representations to CIFAR-10 with random-mask training | ImageNet-100-trained robust encoder versus supervised baseline · CIFAR-10 | top-1 accuracy: 79.83$\pm$0.04 -> 80.93$\pm$0.04 | Helped |
| Transfer the robust representations to STL-10 with random-mask training | ImageNet-100-trained robust encoder versus supervised baseline · STL-10 | top-1 accuracy: 81.09$\pm$0.05 -> 89.44$\pm$0.05 | Helped |
| Transfer the robust representations to ImageNet-100B with random-mask training | ImageNet-100-trained robust encoder versus supervised baseline · ImageNet-100B | top-1 accuracy: 67.45$\pm$0.09 -> 80.09$\pm$0.07 | Helped |
| Transfer the robust representations to COVID X-ray with Gaussian-noise training | ImageNet-100-trained robust encoder versus supervised baseline · COVID X-ray | top-1 accuracy: 83.54$\pm$0.15 -> 79.74$\pm$0.26 | Hurt |
Closest papers by meaning
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.88)
- Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration (AAAI 2024 · similarity 0.87)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.86)
- Exact Posterior Score Estimation for Solving Linear Inverse Problems (NeurIPS 2026 · similarity 0.86)
- FiRe: Fixed-points of Restoration Priors for Solving Inverse Problems (CVPR 2025 · similarity 0.86)
- A Diffusion Model with State Estimation for Degradation-Blind Inverse Imaging (AAAI 2024 · similarity 0.86)
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.