Research map / Probabilistic and scientific modeling / Dynamical systems
Residual Pathway Priors for Soft Equivariance Constraints
Marc Anton Finzi, Gregory Benton, Andrew Gordon Wilson
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 |
|---|---|---|---|
| Use a soft convolutional prior instead of an unconstrained MLP | RPP on CIFAR-10 · CIFAR-10 | classification error: 37.61 ± 0.56 -> 12.62 ± 0.34 | Improved |
| Use a soft convolutional prior instead of a standard convolutional model | RPP on CIFAR-10 · CIFAR-10 | classification error: 12.03 ± 0.46 -> 12.62 ± 0.34 | No clear effect |
| Use a soft convolutional prior instead of an unconstrained MLP | RPP on Energy · Energy | MSE: 0.39 ± 0.48 -> 0.73 ± 0.44 | Mixed |
| Use a soft convolutional prior instead of an unconstrained MLP | RPP on Fertility · Fertility | MSE: 0.049 ± 0.0044 -> 0.060 ± 0.0097 | Hurt |
| Use a soft convolutional prior instead of an unconstrained MLP | RPP on Pendulum · Pendulum | MSE: 4.65 ± 0.50 -> 4.25 ± 0.50 | Improved |
| Use a soft convolutional prior instead of an unconstrained MLP | RPP on Wine · Wine | MSE: 0.66 ± 0.058 -> 0.69 ± 0.031 | Hurt |
| Add the RPP equivariance prior to the transition model | Mujoco transition models · Swimmer-v2 and Hopper-v2 | rollout relative error: Swimmer-v2: 3.9 ± 1.0 -> 2.75 ± \bf{0.31}; Hopper-v2: 9.8±0.5 -> 7.0±\bf{ 0.7} | Improved |
| Add the RPP equivariance prior to the transition model | Mujoco transition models · Ant-v2 | rollout relative error: 16.0\bf{±0.3} -> 20.0±1.1 | Hurt |
| Use RPP-EMLP for both policy and critic rather than policy alone (ablation) | SAC on Mujoco · Mujoco locomotion | average reward | Improved |
| Use RPP-EMLP for policy and critic | SAC on Humanoid-v2 · Humanoid-v2 | training stability | Hurt |
| Use an equivariant architecture instead of RPP when the symmetry is approximate | Dynamical-system models · Modified inertia and windy pendulum | test performance | Hurt |
| Use an equivariant architecture instead of RPP when the specified symmetry is wrong | Dynamical-system models · Misspecified inertia and double pendulum | prediction performance | Hurt |
Closest papers by meaning
- Efficient Reinforcement Learning by Discovering Neural Pathways (NeurIPS 2024 · similarity 0.86)
- A Practical Guide for Incorporating Symmetry in Diffusion Policy (NeurIPS 2025 · similarity 0.86)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025 · similarity 0.85)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.85)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.85)
- Improving Equivariant Model Training via Constraint Relaxation (NeurIPS 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.
