Research map / Architectures and efficiency / Attention and transformers
Flow-based Spectral Kernel Learning for Nonstationary Attention
Zi Yang, Ying Li, Alejandro Lancho, Dario C. Larese, Pablo Olmos, Michael Minyi Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes flow-based spectral kernel attention (FSKA), which learns expressive nonstationary attention kernels to adapt similarity structure beyond fixed or stationary kernels.
Problem
Existing kernelized attention relies on fixed or stationary kernels, which limits its ability to adapt similarity structure to complex token interactions.
Approach
Learns nonstationary attention kernels via a flow-based spectral parameterization, replacing fixed or stationary kernels in kernelized attention with learned ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove query/key and value projections from a standard Transformer (ablation) | Transformer · MNIST | accuracy: 0.8556 -> 0.5267 | Hurt |
| remove query-key projections and fix value projection, with learned flow kernel (ablation) | FSKA · MNIST | accuracy: 0.8656 (IKAN) -> 0.8722 | Improved |
| make value projection learnable (ablation) | FSKA · MNIST | accuracy: 0.8722 -> 0.8725 | No clear effect |
| use positive random features (ELU+1) instead of direct RFF (ablation) | FSKA · MNIST | accuracy: 0.3956 -> 0.8722 | Improved |
| increase RFF dimension M from 32 to 128 (ablation) | FSKA · MNIST | accuracy: 0.8656 -> 0.8822 | Improved |
| increase NGSM mixture count Q from 1 to 10 (ablation) | FSKA · MNIST | accuracy: 0.8483 -> 0.8743 | Improved |
| increase NGSM mixture count Q beyond 10 (ablation) | FSKA · MNIST | accuracy: 0.8743 -> 0.8431 (Q=20) | No clear effect |
| increase flow depth N_f from 2 to 10 (ablation) | FSKA · MNIST | accuracy: 0.8503 -> 0.8707 | Improved |
| replace Transformer attention with FSKA | Transformer · SST-2 | F1-macro: 76.07 -> 78.74 | Improved |
| replace Transformer attention with FSKA | Transformer · SST-5 | F1-macro: 35.11 -> 38.16 | Improved |
| use FSKA instead of Mamba on full documents | Mamba · SCOTUS | accuracy: 59.84 -> 58.91 | Hurt |
| use FSKA instead of Mamba on full documents | Mamba · SCOTUS | macro-F1: 37.91 -> 39.68 | Improved |
Experimental setup
Builds on: kernelized attention
Closest papers by meaning
- The Kernel Reality Check: Benchmarking and Distilling Efficient Attention at Scale (NeurIPS 2026 · similarity 0.88)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.88)
- Implicit Kernel Attention (AAAI 2021 · similarity 0.87)
- Unveiling the Hidden Structure of Self-Attention via Kernel Principal Component Analysis (NeurIPS 2024 · similarity 0.85)
- Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level (NeurIPS 2024 · similarity 0.84)
- Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few (NeurIPS 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.
