Robust Basis Spline Decoupling for the Compression of Transformer Models
A new paper introduces a B-spline-based decoupling framework for compressing transformer models. This method aims to improve numerical stability and expressiveness compared to existing tensor-based decoupling techniques. Experimental results indicate that the proposed approach can significantly reduce parameters while maintaining accuracy in neural network models.
- ▪The paper presents a robust basis spline decoupling framework for transformer model compression.
- ▪It addresses limitations of existing tensor-based decoupling methods by enhancing numerical stability and expressiveness.
- ▪The proposed R-CMTF-BSD algorithm shows promising results in reducing parameters while preserving competitive accuracy.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.18794 (cs) [Submitted on 11 May 2026] Title:Robust Basis Spline Decoupling for the Compression of Transformer Models Authors:Joppe De Jonghe, Van Tien Pham, Mariya Ishteva View a PDF of the paper titled Robust Basis Spline Decoupling for the Compression of Transformer Models, by Joppe De Jonghe and 2 other authors View PDF HTML (experimental) Abstract:Decoupling is a powerful modeling paradigm for representing multivariate functions as compositions of linear transformations and univariate nonlinear functions. A single-layer decoupling can be viewed as a fully connected neural network with a single hidden layer and flexible activation functions, providing a direct link with neural networks.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
