Self supervised learning and foundational models
why is SSL necessary?
- representation learning
- raw data → representations
- generalize across multiple downstream tasks
- compact rep (output dim << input dim)
- robust - invariant to noise
- ex - pretraining AlexNet on ImageNet for img classification, but can be later used for segmentation!
- raw data → representations
visual common sense tasks
contrastive learning
self-distillation
generative reconstruction
beyond images - point clouds, audio, and video
foundational models
applications of SSL
Links:
202609141609