Research map / Image generation / GANs
Joint inference and input optimization in equilibrium networks
Swaminathan Gurumurthy, Shaojie Bai, Zachary Manchester, J Zico Kolter
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Jointly solve the DEQ fixed point and input optimization with an augmented DEQ instead of separate Adam optimization | MDEQ decoder input optimization · CelebA 64×64 | Time: 7659 -> 862 ms | Improved |
| Apply JIIO at test time to the pretrained MDEQ-VAE decoder | MDEQ-VAE · CelebA 64×64 reconstruction | Reconstruction FID: 45.81 -> 42.36 | Improved |
| Train a decoder-only MDEQ with JIIO latent optimization | MDEQ generative model · CelebA 64×64 reconstruction | Reconstruction FID: 45.81 -> 32.52 | Improved |
| Use 100 rather than 40 JIIO iterations (ablation) | JIIO-MDEQ unsupervised inverse solver · CelebA denoising σ=0.2 | Median PSNR: 17.08 -> 17.86 | Improved |
| Use 100 rather than 40 JIIO iterations (ablation) | JIIO-MDEQ unsupervised inverse solver · CelebA denoising σ=0.4 | Median PSNR: 16.03 -> 17.55 | Improved |
| Use 100 rather than 40 JIIO iterations (ablation) | JIIO-MDEQ unsupervised inverse solver · CelebA inpainting | Median PSNR: 15.88 -> 15.87 | No clear effect |
| Replace PGD adversarial training with JIIO adversarial training | MDEQ classifier · MNIST, ε=1, PGD attack | Accuracy: 96.53 ± 0.05 -> 95.74 ± 0.22 | No clear effect |
| Replace PGD adversarial training with JIIO adversarial training | MDEQ classifier · CIFAR, ε=1, PGD attack | Accuracy: 37.4 ± 0.26 -> 37.31 ± 0.67 | No clear effect |
| Replace PGD adversarial training with JIIO adversarial training | MDEQ classifier · CIFAR10, ε=0.5, PGD attack | Accuracy: 51.77 ± 0.75 -> 51.39 ± 0.56 | No clear effect |
| Use JIIO rather than PGD to generate adversarial examples | MDEQ classifier · MNIST adversarial example generation | Time: 4360 -> 1401 ms | Improved |
| Use JIIO rather than Adam for meta-learning inner-loop optimization | DEQ meta-learning model · Omniglot 5-way 1-shot | Time: more than 2.7x lesser time | Improved |
| Use Type-I Anderson rather than Type-II Anderson acceleration (ablation) | JIIO reconstruction optimization · Pretrained MDEQ-VAE decoder | Optimization loss | Mixed |
Closest papers by meaning
- Optimization Induced Equilibrium Networks: An Explicit Optimization Perspective for Understanding Equilibrium Models (TPAMI 2022 · similarity 0.91)
- Deep Equilibrium Models (NeurIPS 2019 · similarity 0.91)
- Neural Deep Equilibrium Solvers (ICLR 2022 · similarity 0.91)
- Separation and Bias of Deep Equilibrium Models on Expressivity and Learning Dynamics (NeurIPS 2024 · similarity 0.91)
- Positive Concave Deep Equilibrium Models (ICML 2024 · similarity 0.88)
- DeltaDEQ: Exploiting Heterogeneous Convergence for Accelerating Deep Equilibrium Iterations (NeurIPS 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.
