Which Features are Learnt by Contrastive Learning? On the Role of Simplicity Bias in Class Collapse and Feature Suppression.
Yihao XueSiddharth JoshiEric GanPin-Yu ChenBaharan MirzasoleimanPublished in: CoRR (2023)
Keyphrases
- selecting features
- image features
- feature vectors
- prior knowledge
- online learning
- feature set
- learning process
- supervised learning
- learning systems
- individual features
- input features
- positive examples
- invariant features
- feature representation
- learning models
- classification accuracy
- positive data
- inductive bias
- binary features
- feature interactions
- class labels
- deep learning
- keypoints
- decision trees
- unsupervised learning
- low level
- reinforcement learning
- feature extraction