Abstract: A new algorithm for feature selection based on information maximization is derived. This algorithm performs subspace mapping from multi-channel signals, where Network Modules (NM) are used to perform the mapping for each of the channels. The algorithm is based on maximizing the Mutual Information (MI) between input and output units of each NM and between output units of different NMs. Such formulation leads to substantial redundancy reduction in output units, in addition to extraction of higher order features from input units that exhibit coherence across time and/or space useful in classification problems. We discuss the performance of the proposed algorithm using two scenarios, one dealing with the classification of EEG data while, the second is a speech application dealing with digit classification.