Research map / Architectures and efficiency / RNNs and normalization
Parallelizing Linear Recurrent Neural Nets Over Sequence Length
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 |
|---|---|---|---|
| Use parallel rather than serial linear recurrence in the kernel | Linear recurrence kernel · Kernel benchmark, sequence length 65,536 | speedup: 38.5; 41.8; 17.5 | Helped |
| Use parallel rather than serial linear recurrence in the kernel | Linear recurrence kernel · Kernel benchmark, sequence length 16 | speedup: 0.06; 0.06; 0.05 | Hurt |
| Use parallel rather than serial linear recurrence | SRU, QRNN, and GILR-LSTM · LS-RNN benchmark, sequence length 65,536 | speedup: 9.21; 6.68; 2.05; 1.41 | Helped |
| Use parallel rather than serial linear recurrence | SRU, QRNN, and GILR-LSTM · LS-RNN benchmark, sequence length 256 | speedup: 0.84; 0.86; 0.99; 0.91 | Hurt |
| Use parallel rather than serial linear recurrence | SRU, QRNN, and GILR-LSTM · LS-RNN benchmark, sequence length 4,096 | speedup: 1.38; 1.18; 1.05; 0.98 | Mixed |
| Use the GILR-LSTM architecture instead of CuDNN LSTM | First-symbol classification · Sequence length 1,024 | wall time to convergence: 0.28 \(\pm\) 0.08 -> 0.031 \(\pm\) 0.002 | Helped |
| Use the GILR-LSTM architecture instead of CuDNN LSTM | First-symbol classification · Sequence length 8,192 | wall time to convergence: 0.58 \(\pm\) 0.06 -> 0.10 \(\pm\) 0.03 | Helped |
| Train GILR-LSTM on the million-step sequence task | GILR-LSTM · First-symbol classification, sequence length 1,048,576 | convergence | Helped |
Closest papers by meaning
- FlashRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models (ICLR 2026 · similarity 0.92)
- Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · similarity 0.91)
- Quasi-Recurrent Neural Networks (ICLR 2017 · similarity 0.90)
- Parallelizing non-linear sequential models over the sequence length (ICLR 2024 · similarity 0.89)
- Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues (ICML 2024 · similarity 0.89)
- Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction (NeurIPS 2026 · 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.