Training models to learn useful features from data with multiple types of input (e.g., images and text).