Distributed learning provides effective tools for big data processing. An effective non-interactive approach for distributed learning is the divide and conquer method. It first partitions a big data set into multiple subsets, then a base algorithm is applied to each subset, and finally the results from these subsets are pooled together. In the context of nonlinear regression analysis, regularized kernel methods usually serve as efficient base algorithms for the second stage. In this talk, I will discuss the minimax optimality of several kernel based regression algorithms in distributed learning.

海报