High-z void galaxies, whose evolution has been driven almost completely free from galaxy mergers, are ideal targets to provide valuable insights into the role of the environment in galaxy evolution. However, a very wide galaxy survey with spectroscopic redshifts are required to find void regions, making it and there have been no studies beyond \(z>3\). In this work, we develop a new deep learning method to select \(z\sim 4\) void galaxies from the g-dropout catalog produced by the HSC-SSP survey; called VoidNet. The VoidNet uses the sky distribution of galaxies and their \((g-r)\) colors as a proxy for redshift despite the large uncertainty to characterize the three-dimensional spatial distribution of galaxies. We train the VoidNet by using Millennium simulation, and when setting a conservative threshold (recall = 0.1%), the VoidNet achieves 90% precision, which is about 20% better than 2D selection. This result shows that deep learning can provide better estimates of the large-scale structure of the universe even when using the photometric data. We are applying this same method to the identification of proto-clusters as well as voids to construct a significantly large sample.