Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition from a Domain Adaptation Perspective

Abstract

Object frequency in the real world often follows a power law, leading to a mismatch between datasets with longtailed class distributions seen by a machine learning model and our expectation of the model to perform well on all classes. We analyze this mismatch from a domain adaptation point of view. First of all, we connect existing classbalanced methods for long-tailed classification to target shift, a well-studied scenario in domain adaptation. The connection reveals that these methods implicitly assume that the training data and test data share the same classconditioned distribution, which does not hold in general and especially for the tail classes. While a head class could contain abundant and diverse training examples that well represent the expected data at inference time, the tail classes are often short of representative training data. To this end, we propose to augment the classic class-balanced learning by explicitly estimating the differences between the class-conditioned distributions with a meta-learning approach. We validate our approach with six benchmark datasets and three loss functions. * Work done while M. Jamal was an intern at Google.

Rethinking the Value of Labels for Improving Class-Imbalanced Learning

“…The result illustrates several interesting aspects. (1) Training data imbalance affects the accuracy of our estimation. For heavily imbalanced training data, we expect the base classifier to have a large difference in accuracy between major and minor classes.…”

“…Literature is rich on learning long-tailed imbalanced data, where the main focuses have been on data re-sampling [2,5,31,41,44], cost-sensitive re-weighting [6,22,23,27,52], as well as class-balanced losses design [7,11,13,26,32]. Other learning paradigms, including transfer learning [33,54], metric learning [55,58], and meta-learning [1,45], have also been explored. Recent studies [25,59] also find that decoupling feature and classifier leads to better long-tailed learning.…”

“…CIFAR-10-LT and CIFAR-100-LT. Following [1,7,11], we use ResNet-32 [20] for all CIFAR-LT experiments. The data augmentation follows [20] to use zero-padding with 4 pixels on each side and then random crop back to the original image size, after which a random horizontal flip is performed.…”

“…iNaturalist 2018. We follow [1,7,11,25] to use ResNet-50 model. Similar to ImageNet-LT, we train all models for 90 epochs in the classifier training stage, and other hyper-parameters are kept the same as in [25].…”

“…Imbalanced data is ubiquitous in the real world, where large-scale datasets often exhibit long-tailed label distributions [1,5,24,33]. In particular, for critical applications related to safety or health, such as autonomous driving and medical diagnosis, the data are by their nature heavily imbalanced.…”