Research map / Architectures and efficiency / Attention and transformers
Residual Shuffle-Exchange Networks for Fast Processing of Long Sequences
Andis Draguns, Emīls Ozoliņš, Agris Šostaks, Matīss Apinis, Karlis Freivalds
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 SE's gated Switch Unit with the residual Switch Unit | Shuffle-Exchange network · LAMBADA test | accuracy: 52.28 -> 54.34 | Improved |
| use RSE instead of the Translation-invariant net | MusicNet · MusicNet | APS: 77.3 -> 78.02 | Improved |
| prepend 1 convolution instead of none (ablation) | RSE · MusicNet, window 1024 | APS: 69.29 -> 69.57 | Improved |
| prepend 2 convolutions instead of none (ablation) | RSE · MusicNet, window 1024 | APS: 69.29 -> 68.95 | No clear effect |
| prepend 3 convolutions instead of none (ablation) | RSE · MusicNet, window 1024 | APS: 69.29 -> 67.39 | Hurt |
| add auxiliary note-prediction losses at regularly spaced sequence positions (ablation) | RSE · MusicNet | APS: 76.84% -> 78.02% | Improved |
| remove LayerNorm from the Residual Switch Unit (ablation) | RSE · binary multiplication, length 128 | test error | Hurt |
| replace GELU with ReLU in the Residual Switch Unit (ablation) | RSE · binary multiplication, length 128 | test error | Hurt |
| remove the residual connection from the Residual Switch Unit (ablation) | RSE · binary multiplication, length 128 | test error | Hurt |
| replace learnable residual scaling with a constant scale of 1 (ablation) | RSE · binary multiplication, length 128 | test error | Hurt |
| increase the Residual Switch Unit hidden layer size (ablation) | RSE · binary multiplication, length 128 | test error | Improved |
Closest papers by meaning
- PoNet: Pooling Network for Efficient Token Mixing in Long Sequences (ICLR 2022 · similarity 0.86)
- Convolutional Sequence to Sequence Learning (ICML 2017 · similarity 0.84)
- Parallelizing Linear Recurrent Neural Nets Over Sequence Length (ICLR 2018 · similarity 0.84)
- Neural Shuffle-Exchange Networks - Sequence Processing in O(n log n) Time (NeurIPS 2019 · similarity 0.83)
- Recurrent Stacking of Layers for Compact Neural Machine Translation Models (AAAI 2019 · similarity 0.83)
- Parallelizing non-linear sequential models over the sequence length (ICLR 2024 · similarity 0.82)
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.
