Instantaneous rain retrieval from passive microwave measurement is a problem that remains challenging even after 30 years of effort from a broad community. The spatial and temporal intermittent nature of rain makes its characterisation very dependent on the considered resolution. By design, passive microwave radiometers (PMR) measure brightness temperature at different resolution according to the considered channel and each channel offers a specific transfer function. Since the mid-90s, Bayesian approaches have proven to be an interesting approach to solve the Tb-to-rain problem. These methods are generally based on a pixel-by-pixel technique providing a probabilistic solution. When coupled with ancillary data (surface type, surface temperature, humidity profile, surface wind) they can give very satisfactory solution with biases close or below 8 to 10%. The natural distribution of rain intensities however, makes the lighter rain much more probable than heavier rain for the same measured brightness temperature set. The only way to discriminate two such situations is often the neighbouring Tbs, indicating if we are, say, in a more convective or a more stratiform area of the rain system. So far, most attempts to take this context into account in the Bayesian schemes have proven weakly successful. New methods based on Machine Learning and Deep-Learning have been developed over the last decade. These are specifically designed to process images as objects and recognize, identify and reconstruct patterns within the objects. It was only natural to try to use these methods on Tbs images and teach them to identify rain patterns. The presentation will show the first results of the application of a U-net type of neural network to retrieve rain from the GPM Microwave Imager (GMI) using a database made of measured Tbs co-located with the Dual Frequency Radar surface rain. Beyond the very encouraging performances, pros and cons of such approaches will also be discussed.