Stochastic Approximation for Canonical Correlation Analysis
share
We study canonical correlation analysis (CCA) as a stochastic optimization problem. We show that regularized CCA is efficiently PAC-learnable. We give stochastic approximation (SA) algorithms that are instances of stochastic mirror descent, which achieve ϵ-suboptimality in the population objective in time poly(1/ϵ,1/δ,d) with probability 1-δ, where d is the input dimensionality.
READ FULL TEXT