Fuzzy Adaptive Focal Loss for Imbalanced Datasets

Tzung-Pei Hong, Wei-Chun Peng, Ja-Hwung Su, Shyue-Liang Wang

Abstract: Artificial intelligence has remarkable effectiveness throughout many domains, and deep neural networks (DNNs) have been shown to outperform other artificial intelligence mechanisms. DNNs imitate how a human brain learns and eliminate the time-consuming process of feature engineering. It can learn implicit information from datasets. However, adequate performance requires large amounts of training data, and real-world applications are often characterized by class imbalance, which results in a bias toward majority classes. In this paper, we use a fuzzy adjustment mechanism to dynamically tune the focal loss hyperparameter based on the three factors: class size, focal loss, and focal loss change. Experimental results on the CIFAR-10 dataset attest to the effectiveness of the proposed method.