Identification of geochemical anomalies is a significant step during regional geochemical exploration. In this matter, new techniques have been developed based on deep learning networks. These simple-structure-networks act like our brains on processing the data by simulating deep layers of thinking. In this paper, a hybrid compositional-deep learning technique was applied to identify the anomalous zones in Dehsalm area which is located in 90 km of SW-Nehbandan, a town in South Khorasan province, Iran. The compositional robust factor analysis (CRFA) was applied as a tool to help select a meaningful subset as an input to Continuous Restricted Boltzmann Machine (CRBM). The dataset consists of 635 stream sediment geochemical samples analyzed for 21 elements. Using CRFA, the 3rd factor (i.e. Pb, Zn, Cu, Ag, Sb, Sr, Ba, Hg and W), indicating epithermal mineralization in the area, was considered as an input set to CRBM. The best-performed CRBM with 80 hidden units and stabilized parameters at 150 iterations was finalized and trained on all the geochemical samples of the study area. Average square contribution (ASC) and average square error (ASE) were determined as anomaly identifiers on the reconstructed error of the trained CRBM. A statistical threshold was applied on the values of the criteria (ASC & ASE) and the resulting outputs were mapped to delineate the anomalous samples. The maps indicated that ASC and ASE have the same performance in the multivariate geochemical anomaly recognition. The anomalies were spatially confirmed with the mineral indexes of Pb, Zn, Cu and Sb, as well as several active mines of Pb and Cu in the study area.