Research map / Image generation / GANs
Task-Aware Compressed Sensing With Generative Adversarial Networks
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 the GAN task-aware for compressed-sensing reconstruction | DCGAN on MNIST · MNIST (m=200, NC=1,000) | per-pixel MSE: 0.0859 -> 0.0322 | Helped |
| train the GAN task-aware for compressed-sensing reconstruction | DCGAN on CelebA · CelebA (m=500, Gaussian A, NC=1,000) | per-pixel MSE: 0.1278 -> 0.0514 | Helped |
| train with compressed measurements only (NC=0) | CSGAN · MNIST and F-MNIST, m=20, 50, 100, 200 | per-pixel MSE | Helped |
| add a compressed-data discriminator while retaining the non-compressed-data discriminator | CSGAN · MNIST and CelebA reconstruction | per-pixel MSE: 0.1138 -> 0.1053; 0.0859 -> 0.0322; 0.0894 -> 0.0124 | Helped |
| use NC=0 rather than NC=100 or NC=1,000 | CSGAN · MNIST (m=200) | per-pixel MSE: 0.0299; 0.1053; 0.0322 | Mixed |
| add contrastive loss to latent z (ablation) | CSGAN · MNIST (m=10) | per-pixel MSE: 0.1042 -> 0.0999 | Helped |
| add contrastive loss to latent z (ablation) | CSGAN · F-MNIST (m=10) | per-pixel MSE: 0.0627 -> 0.0732 | Hurt |
| add contrastive loss to latent z (ablation) | CSGAN · MNIST (m=50, 100, 200, 400) | per-pixel MSE: 0.0353 -> 0.0334; 0.0285 -> 0.0186; 0.0199 -> 0.0139; 0.0169 -> 0.0112 | Helped |
| add contrastive loss to latent z (ablation) | CSGAN · F-MNIST (m=50, 100, 200, 400) | per-pixel MSE: 0.0253 -> 0.0254; 0.0220 -> 0.0203; 0.0179 -> 0.0179; 0.0175 -> 0.0168 | Mixed |
| regularize latent z with contrastive loss (ablation) | CSGAN · F-MNIST classification, LeNet and 50-NN | classification accuracy: 0.4881 vs 0.4372; 0.7410 vs 0.6780; 0.7705 vs 0.7363; 0.7830 vs 0.7584; 0.3937 vs 0.3019; 0.6073 vs 0.4183; 0.6377 vs 0. | Helped |
| use contrastively regularized CSGAN latent features instead of DCGAN latent features | latent-feature classification · MNIST (50-NN, m=39, 78, 196) | classification accuracy: 0.5987 vs 0.3951; 0.6991 vs 0.3957; 0.7656 vs 0.4555 | Helped |
| use contrastively regularized CSGAN latent features rather than DCGAN latent features | latent-feature classification · MNIST (50-NN, m=8) | classification accuracy: 0.3561 vs 0.3679 | Hurt |
Closest papers by meaning
- Deep Compressed Sensing (ICML 2019 · similarity 0.92)
- Compressed Sensing using Generative Models (ICML 2017 · similarity 0.90)
- On the Power of Compressed Sensing with Generative Models (ICML 2020 · similarity 0.89)
- Self-Supervised Generative Adversarial Compression (NeurIPS 2020 · similarity 0.88)
- Invertible generative models for inverse problems: mitigating representation error and dataset bias (ICML 2020 · similarity 0.88)
- Unsupervised Adversarial Image Reconstruction (ICLR 2019 · similarity 0.87)
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.