Fairness in Deep Learning: A Computational Perspective

Fairness of deep learning : Calculated angle

Use deep learning to alleviate discrimination , From three aspects ：Discrimination via Input,Discrimination via Representation,Prediction Quality Disparity Three stages of Pre-processing, In-processing,Post-processing Discuss mitigation methods .

One 、Discrimination via Input

（1）Pre-processing： Replace these fairness sensitive features with substitute values

（2）In-processing： Regularization by model . Regularization implicitly or irregularly optimizes the fairness measure

（3）Post-processing： The prediction and protection properties of the model are used to calibrate the prediction of the model within the reasoning time

Two 、Discrimination via Representation

（1）Pre-processing： Collecting balanced data sets is a possible way to mitigate the representative bias

（2）In-processing： Confrontation training

（3）Post-processing： Suppress neurons that have captured protective properties

3、 ... and 、Prediction Quality Disparity

（1）Pre-processing： Strengthen the diversity of training data sets

（2）In-processing： Standardized model training ----- The migration study