MoCo and BYOL: momentum encoders · MAAI7602 Applied Computer Vision

MoCo keeps a queue of negatives encoded by a slowly updated copy of the network. BYOL drops negatives entirely: an online network with a predictor learns to match a slowly updated target network. How do MoCo and BYOL use a slowly updated copy of the network, with many negatives or with none?

This page needs JavaScript for its explanations, activities and practice questions.