Research map / Probabilistic and scientific modeling / Dynamical systems
Equilibrium Propagation for Non-Conservative Systems
Antonino Emanuele Scurria, Dimitri Vanden Abeele, Bortolo Matteo Mognetti, Serge Massar
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 the antisymmetric-Jacobian correction term to the nudged dynamics (AsymEP vs VF) | feedforward CNN · CIFAR-10 | test accuracy (%): 10.00 -> 89.74 | Improved |
| Use Dyadic EP (doubled state space, energy on augmented space) | feedforward CNN · CIFAR-10 | test accuracy (%): BP 90.66 vs Dyadic EP 90.69 | No clear effect |
| Use AsymEP instead of VF on a feedforward MNIST network | 1 hidden layer with 20 neurons · MNIST | test accuracy (%): 64.3 -> 92.7 | Improved |
| Increase structural asymmetry ratio r_str to 1 under VF (ablation) | VF, 50 hidden neurons · MNIST | test accuracy (%): drops to 5 ± 3 (input-only), 8 ± 4 (all parameters) | Hurt |
| Increase structural asymmetry ratio r_str under AsymEP (ablation) | AsymEP, 50 hidden neurons · MNIST | test accuracy (%): 93.8 -> 94.9 (r_str 0 -> 0.875) | No clear effect |
| Train all parameters including Jdyn rather than only Jin (ablation) | AsymEP · MNIST (fixed asymmetry ratio) | test accuracy (%) | Improved |
| Train Jdyn in addition to Jin under VF (ablation) | VF · MNIST (fixed asymmetry ratio) | collapse point of r_str: collapse at r_str 0.5 -> 0.8 | Improved |
| Use AsymEP with asymmetric connectivity vs EP on symmetric networks | CH dynamics, Fashion-MNIST · Fashion-MNIST (epoch 50) | test accuracy (%): EP symmetric 84.30 -> AsymEP asymmetric 86.78 | Improved |
| Use AsymEP instead of VF on asymmetric and feedforward networks | Fashion-MNIST networks · Fashion-MNIST after 1 epoch | test accuracy (%): on average 15% accuracy gain | Improved |
| Increase depth of AsymEP network from 2 to 3 hidden layers of 500 (ablation) | AsymEP · Fashion-MNIST | test accuracy (%): 86.41 -> 87.8 | Improved |
| Use AsymEP instead of EP and VF from a symmetric initialization | continuous Hopfield network, one hidden layer · MNIST | test accuracy | Improved |
Closest papers by meaning
- Physics-Informed Regularization for Domain-Agnostic Dynamical System Modeling (NeurIPS 2024 · similarity 0.86)
- Improving equilibrium propagation without weight symmetry through Jacobian homeostasis (ICLR 2024 · similarity 0.86)
- Learning Physics Informed Neural ODEs with Partial Measurements (AAAI 2025 · similarity 0.84)
- Semi-supervised Learning of Dynamical Systems with Neural Ordinary Differential Equations: A Teacher-Student Model Approach (AAAI 2024 · similarity 0.84)
- Predicting the Energy Landscape of Stochastic Dynamical System via Physics-informed Self-supervised Learning (ICLR 2025 · similarity 0.84)
- Discovering Physics Laws of Dynamical Systems via Invariant Function Learning (ICML 2025 · similarity 0.84)
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.
