Frame prediction problems, prepare training and validation sets, choose loss functions, and measure generalization. Train simple supervised and unsupervised models before moving to neural generation.
