Research map / Architectures and efficiency / RNNs and normalization
Global Credit Assignment via Dynamical Criticality
Wentao Wang, Keren Gao, Guozhang Chen
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 |
|---|---|---|---|
| replace BPTT with COLA closed-form online rule | vanilla tanh RNN · Adding | test MSE: 0.1426 -> 0.0059 | Improved |
| replace BPTT with COLA | vanilla tanh RNN · Lorenz rollout | MSE: 0.6841 -> 0.0058 | Improved |
| replace BPTT with COLA | vanilla tanh RNN · PTB-Char | loss: 2.8695 -> 1.8352 | Improved |
| replace BPTT with COLA | vanilla tanh RNN · Row-MNIST | accuracy: 0.9713 -> 0.9544 | Hurt |
| replace BPTT with COLA | vanilla tanh RNN · UCI HAR | accuracy: 0.8877 -> 0.8690 | Hurt |
| replace BPTT with COLA | RLIF spiking cell · N-MNIST | accuracy: 98.33% -> 98.54% | Improved |
| replace BPTT with COLA | RadLIF spiking cell · N-MNIST | accuracy: 98.29% -> 93.56% | Hurt |
| replace BPTT with COLA | ConvRNN · refinement and event-frame tasks | accuracy | Hurt |
| use two-point instead of scan-selected gain initialization (ablation) | COLA · Lorenz | MSE: 0.0058 -> 0.0057 | No clear effect |
| use two-point instead of scan-selected gain initialization (ablation) | COLA ConvRNN · Permuted MNIST | accuracy: 0.9136 -> 0.9230 | Improved |
| use near-critical initialization overhead | COLA · 11 benchmarks | one-time overhead: 0.317% median (two-point) | No clear effect |
Closest papers by meaning
- Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues (ICML 2024 · similarity 0.85)
- Retrospective In-Context Learning for Temporal Credit Assignment with Large Language Models (NeurIPS 2025 · similarity 0.84)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.83)
- Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks (ICLR 2024 · similarity 0.83)
- Revisiting Glorot Initialization for Long-Range Linear Recurrences (NeurIPS 2025 · similarity 0.83)
- Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction (NeurIPS 2026 · similarity 0.83)
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.
