In this study, we propose an ensemble learning architecture called "Cognitive Learner", for classification of cognitive states from functional magnetic resonance imaging (fMRI). Proposed architecture consists of a two-layer hierarchy. In the first layer, called voxel layer, we model the connectivity among the voxel time series to represent the detailed information about the experiment. In the second layer, we cluster the voxel time series by using functional similarity measure, to partition the brain volume into homogeneous regions, called super-voxels. Each super-voxel is represented by the average voxel time series that resides in that region. The cognitive states are represented independently in two layers. A set of star meshes are established around each voxel in the first layer and around each super-voxel, in the second layer. The arc weights of the meshes at each layer are estimated by regularized Ridge regression model among the voxels/super-voxels. Mesh arc weights, estimated at voxel and super-voxel levels are used to train independent classifiers. The outputs of two classifiers are ensembled under a stacked generalization architecture. Experiments are carried out to classify the cognitive states in an emotion dataset. Our model achieves 4% higher accuracy on the average compared to state of the art brain decoding models, such as voxel selection methods.