EfficientLearningofSparseRepresentationswithanEnergy-BasedModel