Research map / Optimization / Optimizers
From Perception to Programs: Regularize, Overparameterize, and Amortize
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 |
|---|---|---|---|
| Add min-sampling (3 samples per gradient step) | ROAP · CIFAR-MATH | Program-Acc: 91.8% -> 99.8% | Helped |
| Present images as one contiguous image rather than separate image inputs | ROAP + min-sampling · CIFAR-MATH | Program-Acc: 99.8% -> 99.2% | No clear effect |
| Remove amortized inference (ablation) | ROAP · CIFAR-MATH | Program-Acc: 91.8% -> 28.4% | Hurt |
| Add amortized inference (ablation) | ROAP · 3D reconstruction | Chamfer Distance | No clear effect |
| Remove Gumbel-Softmax (ablation) | ROAP · CIFAR-MATH | Program-Acc: 91.8% -> 21.0% | Hurt |
| Remove Gumbel-Softmax (ablation) | ROAP · 3D reconstruction, Full | Chamfer Distance: 1.7 -> 9.5 | Hurt |
| Use syntax-tree rather than straight-line-code parametrization (ablation) | ROAP · CIFAR-MATH | Program-Acc: 91.8% -> 69.0% | Hurt |
| Reduce maximum program length to 10 lines (ablation) | ROAP · CIFAR-MATH | Program-Acc: 91.8% -> 11.6% | Hurt |
| Replace the symbolic program with a neural network (ablation) | ROAP · 3D reconstruction | Chamfer Distance: Full: 1.7 -> 1.2; Crop-Plane: 1.8 -> 2.0 | Mixed |
| Replace the symbolic program with a neural network (ablation) | ROAP · CIFAR-MATH | Test-Loss: 0.11 -> 0.11 | No clear effect |
| Use syntax-tree rather than straight-line-code parametrization (ablation) | ROAP · 3D reconstruction, Crop-Plane | Chamfer Distance: 1.8 -> 7.1 | Hurt |
| Set program depth to 10 (ablation) | ROAP · 3D reconstruction | Chamfer Distance: Full: 1.7 -> 8.7; Crop-Plane: 1.8 -> 2.7 | Mixed |
Closest papers by meaning
- HOUDINI: Lifelong Learning as Program Synthesis (NeurIPS 2018 · similarity 0.89)
- Safe Neurosymbolic Learning with Differentiable Symbolic Execution (ICLR 2022 · similarity 0.88)
- Bayesian Program Learning by Decompiling Amortized Knowledge (ICML 2024 · similarity 0.88)
- Gradient-Based Program Synthesis with Neurally Interpreted Languages (ICLR 2026 · similarity 0.88)
- Data-Efficient Learning with Neural Programs (NeurIPS 2024 · similarity 0.87)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.87)
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.