Research map / Optimization / Optimizers
Landscape Learning for Neural Network Inversion
Ruoshi Liu, Chengzhi Mao, Purva Tendulkar, Hao Wang, Carl Vondrick
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| optimize in learned space Z via mapping network instead of original latent space | test-time optimization defense · CIFAR-10 adversarial defense (SemiSL, 1 step) | robust accuracy: 34.53 -> 52.11 | Improved |
| optimize in learned space Z via mapping network | StyleGAN inversion · CelebA-HQ (20 steps) | reconstruction loss: 4.456 -> 3.082 | Improved |
| optimize in learned space Z | StyleGAN inversion · LSUN-cat OOD (200 steps) | reconstruction loss: 4.617 -> 3.498 | Improved |
| add coordinate descent training with experience replay buffer (ablation) | mapping network · CelebA-HQ (200 steps) | reconstruction loss: 3.034 -> 1.964 | Improved |
| train mapping network theta instead of random init (ablation) | mapping network · CelebA-HQ (200 steps) | reconstruction loss: 2.723 -> 1.964 | Improved |
| use a randomly initialized untrained mapping network (ablation) | StyleGAN inversion · CelebA-HQ (200 steps) | reconstruction loss: 2.569 -> 2.723 (200 steps), 4.456 -> 4.417 (20 steps) | Mixed |
| use online updates of theta instead of buffered coordinate descent | training algorithm · GAN inversion | convergence rate | Hurt |
| add learned mapping network | VPoser pose reconstruction · GRAB (in-distribution) | final loss: 19% lower than baseline | Improved |
| increase optimization steps from 1 to 5 in learned space | adversarial defense · CIFAR-10 (RO) | robust accuracy: 44.65 -> 43.43 | No clear effect |
Closest papers by meaning
- Invertibility of Convolutional Generative Networks from Partial Measurements (NeurIPS 2018 · similarity 0.87)
- DeepLandscape: Adversarial Modeling of Landscape Videos (ECCV 2020 · similarity 0.87)
- Generative Model Inversion Through the Lens of the Manifold Hypothesis (NeurIPS 2025 · similarity 0.85)
- Sparse Model Inversion: Efficient Inversion of Vision Transformers for Data-Free Applications (ICML 2024 · similarity 0.85)
- Network Inversion of Convolutional Neural Nets (Student Abstract) (AAAI 2025 · similarity 0.85)
- Exploring Landscapes for Better Minima along Valleys (NeurIPS 2025 · similarity 0.84)
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.
