This talk will discuss the role of pretraining as a general paradigm behind the progress in Deep Learning. After introducing the relevant concepts, the talk will highlight the use of pretraining in some of our research work, cutting across diverse domains such as Natural Language Processing and Crystal Property Prediction.