Enhancing few-shot learning using targeted mixup

Irrespective of the attention that long-tailed classification has received over recent years, expectedly, the performance of the tail classes suffers more than the remaining classes. We address this problem by means of a novel data augmentation technique called Targeted Mixup. This is about mixing class samples based on the model's performance regarding each class. Instances of classes that are difficult to distinguish are randomly chosen and linearly interpolated to produce a new sample such that the model can pay attention to those two classes. The expectation is that the model can learn the distinguishing features to improve classification of instances belonging to their respective classes. To prove the efficiency of our proposed methods empirically, we performed experiments using CIFAR-100-LT, Places-LT, and Speech Commands-LT datasets. From the results of the experiments, there was an improvement on the few-shot classes without sacrificing too much of the model performance on the many-shot and medium-shot classes. In fact, there was an increase in the overall accuracy as well.