Decorrelation-Based Deep Learning for Bias Mitigation

Although deep learning has proven to be tremendously successful, the main issue is the dependency of its performance on the quality and quantity of training datasets. Since the quality of data can be affected by biases, a novel deep learning method based on decorrelation is presented in this study. The decorrelation specifically learns bias invariant features by reducing the non-linear statistical dependency between features and bias itself. This makes the deep learning models less prone to biased decisions by addressing data bias issues. We introduce Decorrelated Deep Neural Networks (DcDNN) or Decorrelated Convolutional Neural Networks (DcCNN) and Decorrelated Artificial Neural Networks (DcANN) by applying decorrelation-based optimization to Deep Neural Networks (DNN) and Artificial Neural Networks (ANN), respectively. Previous bias mitigation methods result in a drastic loss in accuracy at the cost of bias reduction. Our study aims to resolve this by controlling how strongly the decorrelation function for bias reduction and loss function for accuracy affect the network objective function. The detailed analysis of the hyperparameter shows that for the optimal value of hyperparameter, our model is capable of maintaining accuracy while being bias invariant. The proposed method is evaluated on several benchmark datasets with different types of biases such as age, gender, and color. Additionally, we test our approach along with traditional approaches to analyze the bias mitigation in deep learning. Using simulated datasets, the results of t-distributed stochastic neighbor embedding (t-SNE) of the proposed model validated the effective removal of bias. An analysis of fairness metrics and accuracy comparisons shows that using our proposed models reduces the biases without compromising accuracy significantly. Furthermore, the comparison of our method with existing methods shows the superior performance of our model in terms of bias mitigation, as well as simplicity of training.