Abstract:

Cluster analysis and and nonlinear dimensionality reduction for visualization are two central machine learning tasks. In this talk I present the approximative learning framework based on large-scale manifolds for both tasks. For clustering, we have used nonnegative matrix factorization based on graph random walk; for visualization we have developed scalable optimization of neighbor embedding. Our methods have significantly improved the state-of-the-art in both tasks. We will also discuss some future work on discovery of macroscopic patterns from big data.