Research map / Image generation / Super-resolution
Controlling Neural Networks via Energy Dissipation
Michael Moeller, Thomas Mollenhoff, Daniel Cremers
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 |
|---|---|---|---|
| constrain network output to be a descent direction with line search, instead of a plain baseline network | DnCNN-style SR network · BSD100 4x super-resolution | PSNR: 26.67 -> 27.12 | Improved |
| constrain network output to be a descent direction with line search, instead of a plain baseline network | DnCNN-style SR network · Set5 4x super-resolution | PSNR: 29.94 -> 31.16 | Improved |
| constrain network output to be a descent direction with line search, instead of a plain baseline network | DnCNN-style SR network · Urban100 4x super-resolution | PSNR: 24.44 -> 24.83 | Improved |
| constrain network output to be a descent direction with line search, instead of a plain baseline network | DnCNN-style SR network · BSD100 4x super-resolution | data-fidelity error: 2.8 · 10^-1 -> 9.8 · 10^-3 | Improved |
| constrain network output to be a descent direction with line search, instead of a plain baseline network | DnCNN-style SR network · Set14 4x super-resolution | SSIM: 0.7642 -> 0.7709 | Improved |
| replace gradient descent with learned energy dissipating descent | gradient descent · BSD100 4x super-resolution | PSNR: 25.09 -> 27.12 | Improved |
| use pretrained DnCNN-B denoising initialization (ablation) | baseline network vs energy dissipating network · 4x super-resolution | PSNR | Mixed |
| replace baseline network with energy dissipating network | Sudoku solver · 50 easy Sudokus | accuracy: 82.6% -> 87.1% | Improved |
| replace baseline network with energy dissipating network | Sudoku solver · 50 easy Sudokus | solve rate: 4.0% -> 52.0% | Improved |
| replace gradient descent with energy dissipating network, 100 iterations | Sudoku solver · 50 easy Sudokus | solve rate: 2.0% -> 52.0% | Improved |
| use lagged updating of the training distribution with the network's own paths (ablation) | energy dissipating network · 2D toy example | qualitative path | Improved |
Closest papers by meaning
- Algorithmic Guarantees for Inverse Imaging with Untrained Network Priors (NeurIPS 2019 · similarity 0.87)
- Parametric Majorization for Data-Driven Energy Minimization Methods (ICCV 2019 · similarity 0.86)
- Scaling Laws For Deep Learning Based Image Reconstruction (ICLR 2023 · similarity 0.86)
- Only Strict Saddles in the Energy Landscape of Predictive Coding Networks? (NeurIPS 2024 · similarity 0.85)
- Weakly Convex Regularisers for Inverse Problems: Convergence of Critical Points and Primal-Dual Optimisation (ICML 2024 · similarity 0.85)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.85)
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.
