Research map / Architectures and efficiency / Attention and transformers
Skyformer: Remodel Self-Attention with Gaussian Kernel and Nystr\"om Method
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 |
|---|---|---|---|
| replace softmax with Gaussian kernel (ablation) | Self-Attention · LRA average | accuracy (%): 57.37 -> 58.56 | Helped |
| replace softmax with Gaussian kernel (ablation) | Self-Attention · LRA Pathfinder | accuracy (%): 65.26 -> 70.73 | Helped |
| replace softmax with Gaussian kernel (ablation) | Self-Attention · LRA Text | accuracy (%): 61.95 -> 60.22 | Hurt |
| replace softmax with Gaussian kernel (ablation) | Self-Attention · LRA running time on Text | running time (h): 4.30 -> 3.91 | No clear effect |
| add Nystrom approximation to Kernelized Attention (Skyformer) (ablation) | Kernelized Attention · LRA Text | running time (h): 3.91 -> 1.02 | Helped |
| add Nystrom approximation to Kernelized Attention (Skyformer) (ablation) | Kernelized Attention · LRA average | accuracy (%): 58.56 -> 59.39 | Helped |
| add Nystrom approximation to Kernelized Attention (Skyformer) (ablation) | Kernelized Attention · LRA Image | accuracy (%): 41.29 -> 40.77 | Hurt |
| use Skyformer instead of self-attention | Self-Attention · LRA Text | accuracy (%): 61.95 -> 64.70 | Helped |
| use Skyformer instead of self-attention | Self-Attention · LRA Retrieval | accuracy (%): 80.69 -> 82.06 | Helped |
| use Skyformer instead of self-attention | Self-Attention · LRA Text | peak memory (GB): 10.37 -> 1.59 | Helped |
| directly apply Nystrom to softmax self-attention (ablation) | Self-Attention · LRA training | training stability | Hurt |
| use Skyformer instead of Nystromformer | Nystromformer · LRA average | accuracy (%): 58.93 -> 59.39 | Helped |
Closest papers by meaning
- Self-Attention through Kernel-Eigen Pair Sparse Variational Gaussian Processes (ICML 2024 · similarity 0.84)
- MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention (NeurIPS 2025 · similarity 0.84)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.84)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.84)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.84)
- Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction (ICLR 2025 · similarity 0.84)
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.