Machine learning has become widespread throughout physical sciences. I will present a pedagogical overview of how neural networks work and how they can be presented as statistical models. After presenting a few recent test cases of how useful machine learning can be for cosmology and astronomy, I will go on to show how and why neural networks themselves are partially flawed for scientific purposes, due to lack of uncertainty quantification. I will then present methods in which we can build neural networks into a statistical framework to regain this trust and understanding.