Switching: understanding the class-reversed sampling in tail sample memorization

Long-tailed visual recognition poses significant challenges to traditional machine learning and emerging deep networks due its inherent class imbalance. Existing reweighting re-sampling methods, although effective, lack a fundamental theory while leaving the paradoxical effects of long tail unsolved, where network failing with head classes under-represented overfitted. In this paper, we investigate long-tailed from memorization-generalization point view, which not only unravels whys previous but also derives new principled solution. Specifically, first empirically identify regularity under distributions, finding that challenge is essentially trade-off between representation high-regularity generalization low-regularity classes. To memorize samples without seriously damaging samples, propose simple yet effective sampling strategy for ordinary mini-batch SGD optimization process, Switching, switches instance-balanced class-reversed once at small rate. By theoretical analysis, show upper bound on error proposed lower than conditionally. our experiments, method can reach feasible performance more efficiently current methods. Further experiments validate superiority Switching strategy, implying could be parsimoniously tackled in memorization stage rate over-exposure samples.