[4G4-OS-26a] OS-26

2:45 PM - 3:00 PM JST(5:45 AM - 6:00 AM UTC)

[4G4-OS-26a-05] Improved Monarch Matrix Approximation Method for AI Model Compression

〇Akihiro Nakamura¹, Keisuke Kawano¹, Naoki Hayashi¹, Tadayoshi Matsumori¹, Takuro Kutsuna¹ (1. Toyota Central R&D Labs., Inc.)

Keywords:

model compression,Monarch matrix,structured matrix,linear layer

As deep learning models continue to scale up, reducing computational resource usage has become increasingly important. In this study, we focus on Monarch matrices, a class of structured matrices that enable efficient computation on GPUs, and reconsider the approximation objective for replacing linear layers in neural networks. While conventional projection-based methods aim to minimize the difference between weight matrices, they are not necessarily optimal from the perspective of output reconstruction. To address this issue, we propose a formulation that directly approximates the output (i.e., the product of the matrix and the input) rather than the weight matrix itself, and solve it using an alternating least squares method. Numerical experiments using randomly generated matrices and inputs demonstrate that the proposed method achieves lower output errors than conventional projection-based approaches.