In the next few years, an international network of gravitational-wave detectors will create a data boom in gravitational-wave data and provide scientists with many opportunities for physics discoveries. On average, every day we will expect to see one gravitational-wave signal and many more false triggers. Therefore, the ability to quickly and reliably extract these signals from noise is extremely important for the future of gravitational-wave and multi-messenger astronomy. In LIGO, noise regression techniques are used to subtract noise artifacts to improve the ability to detect and extract information from gravitational-wave signals. In this talk, I present DeepClean, which is a neural network-based noise regression algorithm that can subtract non-linear and non-stationary noise in the LIGO detector. DeepClean is fast enough to perform in real-time, which can drastically improve the sensitivity to faint, time-sensitive gravitational-wave signals such as neutron star mergers. The framework is generic enough to be applied to other regression problems in other areas of science.