Improved Baselines with Momentum Contrastive Learnins
Moco v2 = MoCo v1 + SimCLR
Moco v1
-
UNsupervised visual representation learning
-
dictionary look-up perspective
\(\rightarrow\) build a dynamic dictionary ( with a queue = FIFO )
-
moving-averaged encoder
SimCLR
- (1) large batch size for lots of negative samples
- (2) stronger augmentation
- (3) MLP projection head
MoCo v2 = Moco V1 + SimCLR (2) & (3)
