Research map / Architectures and efficiency / RNNs and normalization
AntisymmetricRNN: A Dynamical System View on Recurrent Neural Networks
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 |
|---|---|---|---|
| antisymmetric recurrent weight parameterization (with diffusion) replacing unstructured weights (ablation) | RNN on pixel-by-pixel CIFAR-10 · pixel-by-pixel CIFAR-10 | accuracy: 54.6% -> 58.7% | Helped |
| antisymmetric recurrent weight parameterization replacing unstructured weights (ablation) | RNN on noise padded CIFAR-10 · noise padded CIFAR-10 | accuracy: 46.2% -> 48.3% | Helped |
| AntisymmetricRNN vs LSTM on pixel-by-pixel MNIST using about 1/7 of parameters | AntisymmetricRNN · MNIST / permuted MNIST | accuracy: MNIST 97.3% -> 98.0%; pMNIST 92.6% -> 95.8% | Helped |
| input gate added to AntisymmetricRNN | AntisymmetricRNN · permuted MNIST | accuracy: 95.8% -> 93.1% | Hurt |
| input gate added to AntisymmetricRNN | AntisymmetricRNN · pixel-by-pixel MNIST / pixel-by-pixel CIFAR-10 / noise padded CIFAR-10 | accuracy: MNIST 98.0% -> 98.8%; CIFAR pixel 58.7% -> 62.2%; noise padded 48.3% -> 54.7% | Helped |
| AntisymmetricRNN architecture vs LSTM on noise padded CIFAR-10 with long noise padding | AntisymmetricRNN · noise padded CIFAR-10 | accuracy: 11.6% -> 48.3% | Helped |
| AntisymmetricRNN vs LSTM on pixel-by-pixel CIFAR-10 (short-term dependency dominated) | AntisymmetricRNN · pixel-by-pixel CIFAR-10 | accuracy: 59.7% (LSTM) vs 58.7% | No clear effect |
| diffusion constant gamma: very large values (vs small positive values) (ablation) | AntisymmetricRNN · noise padded CIFAR-10 end-to-end Jacobian eigenvalues | eigenvalue mean/variance of end-to-end Jacobian | Hurt |
| diffusion term (subtracting gamma from diagonal) added to antisymmetric weights (ablation) | AntisymmetricRNN forward Euler dynamics · 2D simulation | qualitative trajectory stability | Helped |
Closest papers by meaning
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.90)
- Stable Recurrent Models (ICLR 2019 · similarity 0.90)
- UnICORNN: A recurrent model for learning very long time dependencies (ICML 2021 · similarity 0.90)
- RNNs Incrementally Evolving on an Equilibrium Manifold: A Panacea for Vanishing and Exploding Gradients? (ICLR 2020 · similarity 0.89)
- On the difficulty of learning chaotic dynamics with RNNs (NeurIPS 2022 · similarity 0.89)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.89)
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.