Research map / Architectures and efficiency / Pruning and efficient tuning
Can We Find Strong Lottery Tickets in Generative Models?
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 |
|---|---|---|---|
| Use moment-matching scores to prune a randomly initialized generator to 10% of weights | GFMN · CIFAR-10 | FID: 18.86 -> 15.06 | Helped |
| Use moment-matching scores to prune a randomly initialized generator to 10% of weights | GFMN · FFHQ | FID: 11.52 -> 13.32 | Hurt |
| Use moment-matching scores to prune a randomly initialized generator to 10% of weights | GFMN · LSUN-Bedroom | FID: 17.32 -> 20.21 | Hurt |
| Use moment-matching scores to prune a randomly initialized generator to 10% of weights | GFMN · CelebA | FID: 9.52 -> 10.93 | Hurt |
| Replace per-layer score ranking with global score ranking (ablation) | Edge Popup on randomly initialized SNGAN · CelebA, k=10% | FID: 16.60 -> 15.47 | Helped |
| Replace per-layer score ranking with global score ranking (ablation) | Edge Popup on randomly initialized SNGAN · CelebA, k=20% | FID: 15.99 -> 17.16 | Hurt |
| Replace per-layer score ranking with global score ranking (ablation) | Edge Popup on randomly initialized SNGAN · CelebA, k=30%, 40%, 50%, 60%, 70% | FID: 18.30 -> 17.54; 21.14 -> 16.63; 24.25 -> 14.24; 31.14 -> 12.77; 37.68 -> 21.00 | Helped |
| Keep the last layer unpruned while applying global Edge Popup (ablation) | Pretrained SNGAN · CelebA, k=10% | FID: 14.558 -> 20.253 | Hurt |
| Keep the first and last layers unpruned (ablation) | Pretrained SNGAN with Edge Popup · CelebA, k=10% | FID: 15.309 -> 34.926 | Hurt |
| Use signed Kaiming constant rather than Kaiming normal initialization (ablation) | Randomly initialized generator · GFMN, LSUN-Bedroom | FID | Helped |
| Increase channel width (ablation) | GFMN · CelebA | FID: 26.96; 10.93; 10.12; 8.83; 8.97; 8.31 | Helped |
| Use adversarial loss instead of the moment-matching score (ablation) | Subnetwork search · BigGAN, BabyImageNet | FID | Hurt |
Closest papers by meaning
- Winning Lottery Tickets in Deep Generative Models (AAAI 2021 · similarity 0.93)
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.91)
- Polynomially Over-Parameterized Convolutional Neural Networks Contain Structured Strong Winning Lottery Tickets (NeurIPS 2023 · similarity 0.91)
- Instant Soup: Cheap Pruning Ensembles in A Single Pass Can Draw Lottery Tickets from Large Models (ICML 2023 · similarity 0.91)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.88)
- On the Sparsity of the Strong Lottery Ticket Hypothesis (NeurIPS 2024 · similarity 0.88)
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.