aiqbench analysis
This paper proposes FSKA, which learns nonstationary attention kernels by modeling their spectral density with normalizing flows. It also reports removing query-key projections and fixed orthogonal value projections, cutting projection parameters by roughly 40–47%. The claims rest on controlled classification, ablation, and scaling experiments, so readers should check how broadly the gains hold on larger, real-world benchmarks.
Flow-based Spectral Kernel Learning for Nonstationary Attention
arXiv:2610.30465v1 Announce Type: cross Abstract: Kernelized attention provides a useful perspective for understanding self-attention as a similarity-based aggregation mechanism, but existing approaches often rely on fixed or stationary kernels, limiting their ability to adapt the similarity structure to complex token interactions. In this paper, we propose flow-based spectral kernel attention (FSKA), a framework for learning expressive nonstationary attention kernels. FSKA formulates kernelized attention through spectral density learning and parameterizes the resulting bivariate spectral density with normalizing flows. This design enables flexible and regularizable kernel learning while remaining compatible with scalable linear-attention computation. We further show that expressive kernel learning enables a simplified attention design with removed query--key projections and fixed orthogonal value projections, resulting in a more parameter-efficient architecture. Controlled experiments on classification, ablation, and scalability profiling show that FSKA achieves competitive or improved performance over existing attention mechanisms, while reducing projection-related trainable parameters by approximately $40\%$--$47\%$ compared with most baselines and maintaining favorable linear-attention memory scaling. Code Available: https://github.com/ziyang18/FSKA.
Abstract:Kernelized attention provides a useful perspective for understanding self-attention as a similarity-based aggregation mechanism, but existing approaches often rely on fixed or stationary kernels, limiting their ability to adapt the similarity structure to complex token interactions. In this paper, we propose flow-based spectral kernel attention (FSKA), a framework for learning expressive nonstationary attention kernels. FSKA formulates kernelized attention through spectral density learning and parameterizes the resulting bivariate spectral density with normalizing flows. This design enables flexible and regularizable kernel learning while remaining compatible with scalable linear-attention computation. We further show that expressive kernel learning enables a simplified attention design with removed query--key projections and fixed orthogonal value projections, resulting in a more parameter-efficient architecture. Controlled experiments on classification, ablation, and scalability profiling show that FSKA achieves competitive or improved performance over existing attention mechanisms, while reducing projection-related trainable parameters by approximately $40\%$--$47\%$ compared with most baselines and maintaining favorable linear-attention memory scaling. Code Available: this https URL.
Machine Learning (stat.ML); Machine Learning (cs.LG)
arXiv:2610.30465 [stat.ML]
(or arXiv:2610.30465v1 [stat.ML] for this version)
https://doi.org/10.48550/arXiv.2610.30465 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 254 words