Deep learning has received much attention as of the most powerful approaches for multimodal representation learning in recent years. An ideal model for multimodal data can reason about missing modalities using the available ones, and usually provides more information when multiple modalities are being considered. All the previous deep models contain separate modality-specific networks and find a shared representation on top of those networks. Therefore, they only consider high level interactions between modalities to find a joint representation for them. In this paper, we propose a multimodal deep learning framework (MDLCW) that exploits the cross weights between representation of modalities, and try to gradually learn interactions of the modalities in a deep network manner (from low to high level interactions). Moreover, we theoretically show that considering these interactions provide more intra-modality information, and introduce a multi-stage pre-training method that is based on the properties of multi-modal data. In the proposed framework, as opposed to the existing deep methods for multi-modal data, we try to reconstruct the representation of each modality at a given level, with representation of other modalities in the previous layer. Extensive experimental results show that the proposed model outperforms state-of-the-art information retrieval methods for both image and text queries on the PASCAL-sentence and SUN-Attribute databases.