Decorrelation-Based Deep Learning for Bias Mitigation

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