The decomposition of resting state Functional Magnetic Resonance Imaging (rs-fMRI) data by Dictionary Learning (DL) using sparsity constraint have been recently shown to be an alternative to traditional consideration of independence criteria to obtain Resting State Networks. However, a single level decomposition is not suitable when applied for group rs-fMRI analysis, as it requires a large model order or an equivalently large number of dictionary atoms to appropriately identify distinct brain networks. This is computationally expensive for group rs-fMRI analysis in resource constraint environments. In this work, we have proposed a deep sparse factorization method for multi-level decomposition of rs-fMRI data. Preliminary results from this study show that a multi-layer framework with very small model order can reveal better spatial maps of resting brain that share equivalent temporal dynamics. In addition, the proposed approach outperforms the existing DL approaches with single level decomposition in the metric of quality as well as computational complexity.