Self-Supervised-Self-Labeling-Refinement-with-Bootstrap-Your-Own-Latent

Self Labeling Refinement procedure in Self-Supervised framework Bootstrap Your Own Latent (BYOL)

Introduced two new loss functions for learning robust image representations for doing self-labeling refinement.

1. Cross-cosine Similarity Loss (CCSL)
2. Cross-Sigmoid Similarity Loss (CSSL)

Following are the screenshots from the paper

As we can see that the proposed loss functions, CCSL and CSSL are outperforming the vanilla BYOL setup on STL classification dataset for linear evaluation protocol.