Supervised Exponential Family Principal Component Analysis via Convex Optimization (M39)
Yuhong Guo yuhongguo.cs@gmail.com Problem: supervised dimensionality reduction (n >> t ) Approach: convex supervised exponential family PCA Latent variable model: Z  X , Z  Y Use label information, optimize maxZ log P (X , Y |Z ) Convex formulation & global optimization  avoid the local optima of EM-learning Sample-based approximation to exponential family models  data adaptive model, overcome the limitations of implicit Gaussian assumptions Automatic kernelization achieve nonlinearity