Representation Learning with Statistical Independence to Mitigate Bias

Abstract

Presence of bias (in datasets or tasks) is inarguably one of the most critical challenges in machine learning applications that has alluded to pivotal debates in recent years. Such challenges range from spurious associations between variables in medical studies to the bias of race in gender or face recognition systems. Controlling for all types of biases in the dataset curation stage is cumbersome and sometimes impossible. The alternative is to use the available data and build models incorporating fair representation learning. In this paper, we propose such a model based on adversarial training with two competing objectives to learn features that have (1) maximum discriminative power with respect to the task and (2) minimal statistical mean dependence with the protected (bias) variable(s). Our approach does so by incorporating a new adversarial loss function that encourages a vanished correlation between the bias and the learned features. We apply our method to synthetic data, medical images (containing task bias), and a dataset for gender classification (containing dataset bias). Our results show that the learned features by our method not only result in superior prediction performance but also are unbiased.

Figure 1:

Average face images for each shade category (1^st row), average saliency map of the trained baseline (2^nd row), and BR-Net (3^rd row) color-coded with the normalized saliency for each pixel. BR-Net results in more stable patterns across all 6 shades. The last column shows the tSNE projection of the learned representations by each method. Our method results in a better representation space invariant to the bias variable (shade) while the baseline shows a pattern influenced by the bias. Average accuracy of per-shade gender classification over 5 runs of 5-fold cross-validation (pre-trained on ImageNet, fine-tuned on GS-PPB) is shown on each average map. BR-Net not only obtains better accuracy for the darker shade but also regularizes the model to improve overall per-category accuracy.

Figure 2:

BR-Net architecture: $\mathbb{F}\mathbb{E}$ learns features, F, that successfully classify ($\mathbb{C}$) the input while being invariant (statistically independent) to the protected variables, b, using $\mathbb{B}\mathbb{P}$ and the adversarial loss, −λL_bp (based on correlation coefficient). Forward arrows show forward paths while the backward dashed ones indicate back-propagation with the respective gradient (∂) values.

Figure 3:

BR-Net can remove direct dependency between F and b for both dataset or task bias.

Figure 4:

Formation of synthetic dataset (a) and comparison of results for different methods (b).

Figure 5:

tSNE projection of the learned features for different methods. Color indicates the value of σ_B.

Figure 6:

Statistical dependence between the learned features and age for the CTRL cohort in the HIV experiment, which is quantitatively measured by dcor².

Figure 7:

Accuracy, TNR, and TPR of the HIV experiment, as a function of the # of iterations for (a) 3D CNN baseline, (b) BR-Net. Our method is robust against the imbalanced age distribution between HIV and CTRL.

Figure 8:

Accuracy of gender prediction from face images across all shades (1 to 6) of the GS-PPB dataset with two backbones, (left) VGG16 and (right) ResNet50. BR-Net consistently results in more accurate predictions in all 6 shade categories.

Figure 9:

Learned representations by different methods. Color encodes the 6 categories of skin shade.