We characterise and quantify the representation learning disparity that emerges at 430:1 class imbalance — far beyond the ~100:1 ratio where standard SSL robustness holds. Under uniform sampling, ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results