The Unreasonable Effectiveness of the Class-reversed Sampling in Tail Sample Memorization

Long-tailed visual recognition poses significant challenges to traditional machine learning and emerging deep networks due to its inherent class imbalance. A common belief is that tail classes with few samples cannot exhibit enough regularity for pattern extraction. What makes things worse, the limited cardinality may lead to low exposure of tail classes in the training stage. Re-sampling methods, especially those who naively enlarge the exposure frequency, eventually fail with head classes under-represented and tail classes overfitted. Arguing that long-tailed learning involves a trade-off between head class pattern extraction and tail class memorizing, we first empirically identify the regularity of classes under long-tailed distributions and find that regularity of the same training samples will be sharply decreased with the reduction of class cardinality. Motivated by the recent success of a series works on the memorization-generalization mechanism, we propose a simple yet effective training strategy by switching from instance-balanced sampling to class-reversed sampling to memorize tail classes without seriously damaging the representation of head classes. Closely after- wards, we give the theoretical generalization error upper bound to prove that class- reversed sampling is better than instance-balanced sampling during the last train- ing stage. In our experiments, the proposed method can reach the state-of-the-art performance more efficiently than current methods, on several datasets. Further experiments also validate the superior performance of the proposed sampling strategy, implying that the long-tailed learning trade-off could be effectively tackled only in the memorization stage with a small learning rate and over-exposure of tail samples.