Why Regularized Auto-Encoders learn Sparse Representation?