A gridded ocean subsurface salinity dataset with global coverage is useful for research on climate change and its variability. Here, we explore the feed-forward neural network (FFNN) approach to reconstruct a high-resolution (0.25 degrees x 0.25 degrees) ocean subsurface (1-2000 m) salinity dataset for the period 1993-2018 by merging in situ salinity profile observations with high-resolution (0.25 degrees x 0.25 degrees) satellite remote-sensing altimetry absolute dynamic topography (ADT), sea surface temperature (SST), sea surface wind (SSW) field data, and a coarse-resolution (1 degrees x 1 degrees) gridded salinity product. We show that the FFNN can effectively transfer small-scale spatial variations in ADT, SST, and SSW fields into the 0.25 degrees x 0.25 degrees salinity field. The root-mean-square error (RMSE) can be reduced by -11 % on a global-average basis compared with the 1 degrees x 1 degrees salinity gridded field. The reduction in RMSE is much larger in the upper ocean than the deep ocean because of stronger mesoscale variations in the upper layers. In addition, the new 0.25 degrees x 0.25 degrees reconstruction shows more realistic spatial signals in the regions with strong mesoscale variations, e.g., the Gulf Stream, Kuroshio, and Antarctic Circumpolar Current regions, than the 1 degrees x 1 degrees resolution product, indicating the efficiency of the machine learning approach in bringing satellite observations together with in situ observations. The large-scale salinity patterns from 0.25 degrees x 0.25 degrees data are consistent with the 1 degrees x 1 degrees gridded salinity field, suggesting the persistence of the large-scale signals in the high-resolution reconstruction. The successful application of machine learning in this study provides an alternative approach for ocean and climate data reconstruction that can complement the existing data assimilation and objective analysis methods.