De-Biasing Methods in Neural Networks: A Survey
Yixin Wang, Han Liu · 2023
Bias is a common problem in both human cognition and machine learning tasks. However, machines struggle more than humans with bias reduction, mainly because most algorithms rely on the assumption that the training data represent a full population in the real world. This assumption does not hold in most cases, i.e., collecting unbiased or representative data is challenging due to various factors such as sampling errors, human prejudices, and cultural differences. Therefore, bias mitigation has become an essential task in artifical intelligence. In recent years, significant progress has been made in improving the effectivess of debiasing, especially in computer vision. This paper provides a comprehensive review of de-biasing methods for neural networks trained on image data. Firstly, we present a formal defination of the bias mitigation problem and discuss some relevant topics. Moreover, we categorize the existing methods into three main types, namely, data-level methods that aim to balance or augment the training data, model-level methods that aim to modify or regularize the learning algorithms, and adversarial and ensembling approaches that use additional models to capture the biases present in data and to guide the training of the primary model. Finally, we conclude this survey and suggest potential research directions for the future.