🍀Lecture 26-27

● Self-supervised learning

Purpose：不用human label

On downstream（have label

○ Pretext then transfer

● Context prediction（内容预测）, Colorization, Rotation prediction

Know category

● Contrastive learning（建设性的，推断

Pair /set of patch；

● Masked Autoencoder

Predict masked value

Know sematics

● Multi-modal self-supervised learning

Know you muti

Eg audio