Image fairness in deep learning: problems, models, and challenges

Abstract

In recent years, it has been revealed that machine learning models can produce discriminatory predictions. Hence, fairness protection has come to play a pivotal role in machine learning. In the past, most studies on fairness protection have used traditional machine learning methods to enforce fairness. However, these studies focus on low dimensional inputs, such as numerical inputs, whereas more recent deep learning technologies have encouraged fairness protection with image inputs through deep model methods. These approaches involve various object functions and structural designs that break the spurious correlations between targets and sensitive features. With these connections broken, we are left with fairer predictions. To better understand the proposed methods and encourage further development in the field, this paper summarizes fairness protection methods in terms of three aspects: the problem settings, the models, and the challenges. Through this survey, we hope to reveal research trends in the field, discover the fundamentals of enforcing fairness, and summarize the main challenges to producing fairer models.

Explainable deep learning in plant phenotyping

“…Arrieta et al ( 2020 ) reviewed 400 articles and proposed a novel definition of explainability, and emphasized that XAI is necessary to ensure security. In machine learning, fairness is considered a subsection of machine learning interpretability and addresses the social and ethical consequences of machine learning algorithms (Tian et al, 2022 ). Linardatos et al ( 2020 ) studied the fairness of machine learning models where the authors mentioned that researchers favor groups of individuals with different attributes over ensuring individuals are treated similarly; thus, the importance of individuals is often ignored.…”

Lung-Centric Feature Analysis for Accurate Pneumonia Detection in Chest X-RAY Images

“…Different combinations of hyperparameters were systematically searched and evaluated to find the best configuration for each model. The GridSearch [47] was utilized to identify the optimal hyperparameters for our models. The performance of each configuration was evaluated using the F1 score as the scoring metric.…”

“…While deep learning techniques are powerful in image classification tasks, their applicability is constrained due to their requisites of substantial datasets and considerable computational resources. This may limit the scalability and accessibility of the models for some users or scenarios [16,47]. For the hand-crafted-based classifiers, the literature has a common drawback in which the features are extracted from the entire X-ray images.…”

Drop the shortcuts: image augmentation improves fairness and decreases AI detection of race and other demographics from medical images

“… 29 , 33 Multiple studies have demonstrated that data augmentation can effectively eliminate learned shortcuts from the original dataset. 34 , 35 , 36 This is further evidenced by a recent study employing an adversarial U-Net architecture to alter natural images, thereby removing shortcut features. 36 If shortcut learning potentiates bias in healthcare DL algorithms, data augmentation may assist in improving model fairness by counteracting shortcut learning.…”