Research map / Optimization / Optimizers
Optimization Induced Equilibrium Networks: An Explicit Optimization Perspective for Understanding Equilibrium Models
Xingyu Xie, Qiuhao Wang, Zenan Ling, Xia Li, Guangcan Liu, Zhouchen Lin
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 |
|---|---|---|---|
| add decorrelation regularization R*z (alpha=0.4) (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 86.9% -> 88.0% | Improved |
| add decorrelation regularization R-dagger z (alpha=0.01) (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 58.4% -> 72.7% | Improved |
| add decorrelation R-dagger z (alpha=0.4) (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 86.9% -> 86.5% | Hurt |
| add HSIC regularizer (R*z) | deep OptEq 162k · CIFAR-10 | accuracy: 85.7% -> 87.4% | Improved |
| use deep OptEq with 5 layers instead of a general DNN (ablation) | DNN · CIFAR-10 (199k params) | accuracy: 82.7% -> 88.0% | Improved |
| increase alpha from 0.1 to 0.4 (ablation) | deep OptEq no Reg · CIFAR-10 (199k params) | accuracy: 56.8% -> 86.9% | Improved |
| increase unrolling iterations K from 5 to 20 (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 83.52% -> 87.71% | Improved |
| increase unrolling iterations K from 20 to 40 (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 87.71% -> 87.83% | No clear effect |
| use unrolling (K=20) instead of IFT (thd=1e-03) training (ablation) | deep OptEq · CIFAR-10 (199k params) | accuracy: 87.63% -> 87.71% | Improved |
| loosen IFT residual threshold from 1e-03 to 1e-02 (ablation) | IFT deep OptEq · CIFAR-10 (199k params) | accuracy: 87.63% -> 85.95% | Hurt |
| use deep OptEq backbone instead of FCN | segmentation network · Cityscapes val | mIoU: 71.47 -> 74.47 | Improved |
| use deep OptEq instead of Aug. Neural ODE / MON | implicit models · CIFAR-10 | accuracy: 60.6% -> 85.7% | Improved |
Closest papers by meaning
- Joint inference and input optimization in equilibrium networks (NeurIPS 2021 · similarity 0.91)
- Positive Concave Deep Equilibrium Models (ICML 2024 · similarity 0.90)
- Separation and Bias of Deep Equilibrium Models on Expressivity and Learning Dynamics (NeurIPS 2024 · similarity 0.90)
- Subhomogeneous Deep Equilibrium Models (ICML 2024 · similarity 0.88)
- DeltaDEQ: Exploiting Heterogeneous Convergence for Accelerating Deep Equilibrium Iterations (NeurIPS 2024 · similarity 0.87)
- Understanding Representation of Deep Equilibrium Models from Neural Collapse Perspective (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.
