3 papers across 2 sessions
We introduce a regularizer that promotes diverse, task-relevant features over spurious ones in contrastive learning.