The Role of Augmentations in Perception and RL

The representation network in policy learning can benefit from supervised learning with data augmentation./nAugmentations are useful for perception, but may not be as effective for RL./nData augmentation can improve sample efficiency in RL, but may not improve generalization significantly./nIncorporating an auxiliary objective with data augmentation can make optimization easier./nAdding data augmentation to a network that does not fully fit the data can increase variance and affect Q target values.