Dynamic Error-Bounded Hierarchical Matrices in Neural Network Compression

Abstract

This paper presents an innovative framework that integrates hierarchical matrix (H-matrix) compression techniques into the structure and training of Physics-Informed Neural Networks (PINNs). By leveraging the low-rank properties of matrix sub-blocks, the proposed dynamic, error-bounded H-matrix compression method significantly reduces computational complexity and storage requirements without compromising accuracy. This approach is rigorously compared to traditional compression techniques, such as Singular Value Decomposition (SVD), pruning, and quantization, demonstrating superior performance, particularly in maintaining the Neural Tangent Kernel (NTK) properties critical for the stability and convergence of neural networks. The findings reveal that H-matrix compression not only enhances training efficiency but also ensures the scalability and robustness of PINNs for complex, large-scale applications in physics-based modeling. This work offers a substantial contribution to the optimization of deep learning models, paving the way for more efficient and practical implementations of PINNs in real-world scenarios.