Community detection has become pervasive in finding similar patterns present in the network. It aims to discover lower dimensional embedding for representing the structure of network. Many real-life networks comprise overlapping communities and have non-linear features. Despite of having a great potential in analyzing the network structure, the existing approaches provide a limited support and find disjoint communities only. As data is growing unprecedentedly, scalable and intelligent solutions are obligatory for identifying similar patterns. Motivated by the robust representation ability of deep neural network based autoencoder, we proposed a learning model named ‘DeCom’ for finding overlapping communities from large networks. DeCom uses autoencoder based layered approach to initialize candidate seed nodes and to determine the number of communities by considering the network structure. The selected seed nodes and formed clusters are refined in last layer by minimizing the reconstruction error using modularity. The performance of DeCom is compared with three state-of-art clustering algorithms by using real life networks. It is observed that the felicitous selection of seed nodes reduces the number of iterations. The experimental results reveal that the proposed DeCom scales up linearly to handle large graphs and produces better quality of clusters when compared with the other state-of-art clustering algorithms.