The cosmic 21-cm signal is a powerful probe into the first billion years of the cosmic evolution. Its power in constraining cosmology and astrophysics of these early times comes from the fact it covers half of the observable Universe. This imposes hard challenges for the data analysis of the future radiotelescope surveys. In this talk, we will discuss several ways in which machine learning plays an important role for the inference from the cosmic 21-cm. We give an insight into questions such as: “How to optimally encode the observed 21-cm signal?” - where we explored recurrent NNs for the task, “What is the likelihood of the 21-cm signal for a fixed compression?” - where simulation-based inference offers a solution, and “How to quantify the quality of different summaries of the signal?” - where we introduced Fisher-based metric for the task.