The past five years have seen rapid progress in large-scale pre-trained models across a variety of domains, such as computer vision, natural language processing, robotics, bioinformatics, etc. Leveraging a huge number of parameters, large-scale pre-trained models are capable of encoding rich knowledge from labeled and/or unlabeled examples. Supervised and self-supervised pre-training have been the two most representative paradigms, through which pre-trained models have demonstrated large benefits on a wide spectrum of downstream tasks. There are also other pre-training paradigms, e.g., meta-learning for few-shot learning, where pre-trained models are trained so that they quickly adapt to solve new tasks. However, there are still many remaining challenges and new opportunities ahead for pre-training, In this workshop, we propose to have the following two foci: (1) Which pre-training methods transfer across different applications/domains, which ones don't, and why? (2) In what settings should we expect pre-training to be effective, compared to learning from scratch?