SEER: large-scale SSL for vision.
— Yann LeCun (@ylecun) March 4, 2021
- pre-train via SSL on 1 billion randomly selected images using SwAV.
- fine-tune on ImageNet: 84.2% top-1 accuracy.
- ft on 10% of ImageNet: 77.9%
- ft on 1% (13 samples per class): 60.5%
- beats SOTA on other CV taskshttps://t.co/Q8BT5QvKmf