Machine learning (ML) and statistical signal processing provide a powerful lens through which to develop and understand new imaging techniques. Together they allow one to abstract complex physical systems into manageable representations that can leverage new kinds of models and algorithms, such as deep learning. When used appropriately, ML-based imaging systems enable a host of new capabilities, from imaging around corners to imaging through tissue and fog. These advancements have wide-sweeping implications in scientific imaging, medical imaging, consumer photography, navigation, security, and more. The key to successfully applying ML to imaging is to carefully incorporate accurate physical models and statistics. In this talk, I will describe how physical models and statistics enable ML-based imaging without training data, ML-based optical system design, and ML-based imaging around corners and through keyholes.