The paper introduces a new framework for feature learning in classification motivated by information theory. We first systematically study the information structure and present a novel perspective revealing the two key factors in information utilization: class-relevance and redundancy. We derive a new information decomposition model where a novel concept called class-relevant redundancy is introduced. Subsequently a new algorithm called Conditional Informative Feature Extraction is formulated, which maximizes the joint class-relevant information by explicitly reducing the class-relevant redundancies among features. To address the computational difficulties in information-based optimization, we incorporate Parzen window estimation into the discrete approximation of the objective function and propose a Local Active Region method which substantially increases the optimization efficiency. To effectively utilize the extracted feature set, we propose a Bayesian MAP formulation for feature fusion, which unifies Laplacian Sparse Prior and Multivariate Logistic Regression to learn a fusion rule with good generalization capability. Realizing the inefficiency caused by separate treatment of the extraction stage and the fusion stage, we further develop an improved design of the framework to coordinate the two stages by introducing a feedback from the fusion stage to the extraction stage, which significantly enhances the learning efficiency. The results of the comparative experiments show remarkable improvements achieved by our framework.