Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance

Class-imbalance is one of the major challenges in real world datasets, where a few classes (called majority classes) constitute much more data samples than rest minority classes). Learning deep neural networks using such datasets leads to performances that are typically biased towards classes. Most prior works try solve class-imbalance by assigning weights various manners (e.g., re-sampling, cost-sensitive learning). However, we argue number available training may not be always good clue determine weighting strategy because some might sufficiently represented even small data. Overweighting can lead drop model’s overall performance. We claim ‘difficulty’ class as perceived model important weighting. In this light, propose novel loss function named Class-wise Difficulty-Balanced loss, or CDB which dynamically distributes each sample according difficulty belongs to. Note assigned change for with learning progress. Extensive experiments conducted on both image (artificially induced class-imbalanced MNIST, long-tailed CIFAR and ImageNet-LT) video (EGTEA) datasets. The results show consistently outperforms recently proposed functions irrespective type (i.e., image).