Department of Mathematics - Seminar on Applied Mathematics - Data-Driven Matrix Computations for Kernel Methods
Supporting the below United Nations Sustainable Development Goals:支持以下聯合國可持續發展目標:支持以下联合国可持续发展目标:
Dense kernel matrices arise in a broad range of disciplines, such as potential theory, molecular biology, statistical machine learning, etc. To reduce the computational cost, low-rank or hierarchical low-rank techniques are often used to construct an economical approximation to the original matrix. In this talk, we propose data-driven approaches for accelerating dense kernel matrix computations. We first provide a straightforward geometric interpretation that answers a central question: what kind of subset is preferable for skeleton low-rank approximations. Based on the geometric findings, we present scalable and robust hierarchical algorithms for black-box dense kernel matrix computations. The efficiency and robustness will be demonstrated through experiments for various datasets, kernels, and dimensions, including benchmark comparison to the state-of-the-art packages for N-body simulations.