Recent progress in deep learning methods for the task of source separation have significantly advanced the state-of-the-art. Among all the recent proposals, end-to-end systems that take waveform as input and directly generate waveforms have shown their advantage on both the system performance and the flexibility. In this talk, I will briefly go through some of the recent advances in the problem of end-to-end neural source separation. I will start with the general problem definition of source separation, then introduce several single-channel and multi-channel approaches, and conclude with the challenges and future works in this area.