Large-scale Spectral Clustering for GPU-based Platforms
Guanlin He

24 March 2020, 10:30 Salle/Bat : 465/PCRI-N
Contact :

Activités de recherche : High-performance computing

Résumé :

Clustering is one of the most important tasks in machine learning and data analysis. It aims at exploring the intrinsic structure of data by grouping them into meaningful classes in an unsupervised way. Based on algebraic graph theory, spectral clustering has attracted extensive attention for its fundamental advantages (e.g. global high-quality solution, able to discover arbitrary shaped clusters) compared to traditional clustering algorithms (e.g., k-means). However, spectral clustering has a high-order computational complexity O(n^3) (where n is the number of data instances), especially for eigenvector computations, which becomes an obstacle for its generalization to large-scale applications. This talk will give a global and structured view of the state-of-art of spectral clustering and propose some ideas of parallelization by leveraging GPU-based massively parallel architectures to address large problems.