Misclassification-guided loss under the weighted cross-entropy loss framework

Abstract

As deep neural networks for visual recognition gain momentum, many studies have modified the loss function to improve the classification performance on long-tailed data. Typical and effective improvement strategies are to assign different weights to different classes or samples, yielding a series of cost-sensitive re-weighting cross-entropy losses. Granted, most of these strategies only focus on the properties of the training data, such as the data distribution and the samples’ distinguishability. This paper works these strategies into a weighted cross-entropy loss framework with a simple production form (\(\text {WCEL}_{\prod }\)), which takes into account different features of different losses. Also, there is this new loss function, misclassification-guided loss (MGL), that generalizes the class-wise difficulty-balanced loss and utilizes the misclassification rate on validation data to update class weights during training. In respect of MGL, a series of weighting functions with different relative preferences are introduced. Both softmax MGL and sigmoid MGL are derived to address the multi-class and multi-label classification problems. Experiments are undertaken on four public datasets, namely MNIST-LT, CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and a self-built dataset of 4 main-classes, 44 sub-classes, and altogether 57,944 images, where the results show that on the self-built dataset, the exponential weighting function achieves higher balanced accuracy than the polynomial function does. Ablation studies also show that MGL sees better performance in combination with most of other state-of-the-art loss functions under the \(\text {WCEL}_{\prod }\) framework.