Research map / Architectures and efficiency / Pruning and efficient tuning
Winning Lottery Tickets in Deep 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 |
|---|---|---|---|
| Find a sparse subnetwork with iterative magnitude pruning and late rewinding | Deep generative models · MNIST, CIFAR-10, Celeb-A | sparsity at which performance is preserved: 89.2%; 95.6%; 98.5% | Helped |
| Find a sparse subnetwork with iterative magnitude pruning and late rewinding | VAEs · CIFAR-10, Celeb-A | sparsity at which FID is preserved: 79%; 86.5%; 93.1% | Helped |
| Find a sparse subnetwork with iterative magnitude pruning and late rewinding | GANs · CIFAR-10, Celeb-A | sparsity at which FID is preserved: 83.2%; 89.2%; 79%; 73.7%; 59% | Helped |
| Use winning-ticket initializations instead of random initializations | Generative-model tickets · CIFAR-10, Celeb-A, MNIST | FID, reconstruction loss, or discriminator loss | Helped |
| Prune both components rather than only one component (ablation) | AutoEncoders · MNIST, CIFAR-10, Celeb-A | reconstruction loss | Helped |
| Prune encoder and decoder at comparable levels rather than pruning only one component (ablation) | GANs · CIFAR-10, Celeb-A | FID | Helped |
| Prune the encoder while leaving the decoder unpruned (ablation) | VAE · CIFAR-10, Celeb-A | FID | No clear effect |
| Prune the decoder while leaving the encoder unpruned (ablation) | VAE · CIFAR-10, Celeb-A | FID | Hurt |
| Transfer the VAE decoder winning ticket to the DCGAN generator (ablation) | DCGAN · CIFAR-10 | FID | No clear effect |
| Transfer the GAN generator winning ticket to the VAE decoder (ablation) | VAE · CIFAR-10 | FID | No clear effect |
| Use channel-pruned early-bird tickets instead of an unpruned network | DCGAN · CIFAR-10 | FLOPs: 1.12e+16 -> 0.13e+16 | Helped |
| Use channel-pruned early-bird tickets instead of an unpruned network | DCGAN · CIFAR-10 | Training Time (seconds): 4590.5 -> 2417.07 | Helped |
Closest papers by meaning
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.93)
- Can We Find Strong Lottery Tickets in Generative Models? (AAAI 2023 · similarity 0.93)
- Plant 'n' Seek: Can You Find the Winning Ticket? (ICLR 2022 · similarity 0.93)
- One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers (NeurIPS 2019 · similarity 0.92)
- Why Lottery Ticket Wins? A Theoretical Perspective of Sample Complexity on Sparse Neural Networks (NeurIPS 2021 · similarity 0.92)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · similarity 0.91)
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.