With limited statistics and spatial resolution of current detectors, accurately localizing and separating -ray point sources from the dominating interstellar emission in the GeV energy range is challenging. Motivated by the challenges of the traditional methods used for the -ray source detection, here we demonstrate the application of deep learning based algorithms to automatically detect and classify point sources, which can be applied directly to the binned Fermi-LAT data and potentially be generalized to other wavelengths. For the point source detection task, we use popular deep neural network structure, U-NET and, together with image segmentation, for precise localization of sources, various clustering algorithms were tested on the segmented images. The training samples are based on the source properties of AGNs and PSRs from the latest Fermi-LAT source catalog in addition to the background interstellar emission. Results obtained using these algorithms are shown to be comparable with the traditional methods and checked the robustness by using different interstellar emission models. We also publish our training data-sets and analysis scripts and, invite the community to a challenge -find the best possible method for localizing and classifying -ray sources. Finally, we present a more complex but robust training data generation exploiting full detector potential, which will be used in our follow-up work.