Aviv Regev

For datasets with high-magnitude noise features, joint-embedding is more robust than reconstruction for self-supervised learning.