Title:A Simple Framework for Contrastive Learning of Visual Representations
Author:Ting Chen, Geoffrey Hinton... (Google Research)
參考:Hinton組力作:ImageNet無監督學習最佳性能一次提升7%,媲美監督學習
網絡結構
Data augmentation
2. No single transformation suffices to learn good representations
3. it is critical to compose cropping with color distortion
4. data augmentation that does not yield accuracy benefits for supervised learning can still help considerably with
contrastive learning.
Architectures for Encoder and Head
1. Unsupervised contrastive learning benefits (more) from bigger models
2. G(*):A nonlinear projection is better than a linear projection
3. Contrastive learning benefits (more) from larger batch sizes and longer training