Research map / Architectures and efficiency / Attention and transformers
Neural Shuffle-Exchange Networks - Sequence Processing in O(n log n) Time
Karlis Freivalds, Emīls Ozoliņš, Agris Šostaks
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 the Shuffle-Exchange network instead of the Gated-Attention Reader | LAMBADA word prediction · LAMBADA test | accuracy (%): 49.0 -> 52.28 | Improved |
| Use the Shuffle-Exchange network instead of the Universal Transformer | LAMBADA word prediction · LAMBADA test | accuracy (%): 56.0 -> 52.28 | Hurt |
| Generalize the learned algorithm to sequences longer than the training examples | Shuffle-Exchange network · Duplication and reversal, length 512 | accuracy | Improved |
| Generalize the learned algorithm to sequences longer than the training examples | Shuffle-Exchange network · Multiplication | accuracy | Hurt |
| Add the fixed swapHalf operation rather than use identity (ablation) | Shuffle-Exchange Switch Unit · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Add residual connections between Beneš blocks (ablation) | Shuffle-Exchange network · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Use two shuffle directions as in the Beneš network (ablation) | Shuffle-Exchange network · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Use fixed swapHalf instead of an additional swap gate (ablation) | Shuffle-Exchange Switch Unit · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Keep the Switch Unit instead of replacing it with two fully connected layers (ablation) | Shuffle-Exchange network · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Keep the reset-gate portion instead of replacing it with two fully connected layers while retaining the update gate (ablation) | Shuffle-Exchange Switch Unit · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Retain the update gate (ablation) | Shuffle-Exchange Switch Unit · Multiplication, LAMBADA, addition, and sorting | test accuracy | Improved |
| Increase Beneš block count (ablation) | Shuffle-Exchange network · Multiplication and LAMBADA | test accuracy | Improved |
Closest papers by meaning
- Residual Shuffle-Exchange Networks for Fast Processing of Long Sequences (AAAI 2021 · similarity 0.83)
- Learning and processing the ordinal information of temporal sequences in recurrent neural circuits (NeurIPS 2023 · similarity 0.80)
- Sequence Modeling with Spectral Mean Flows (NeurIPS 2025 · similarity 0.79)
- Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks (ICLR 2024 · similarity 0.77)
- Interacting Diffusion Processes for Event Sequence Forecasting (ICML 2024 · similarity 0.77)
- Fundamental limits of learning in sequence multi-index models and deep attention networks: high-dimensional asymptotics and sharp thresholds (ICML 2025 · similarity 0.76)
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.
