GAN; A PROMISING APPROACH TO MITIGATE THE PROBLEM OF BIAS IN AI (original) (raw)

2021, The Society for Philosophy and Technology Conference (SPT)

The fact that deep neural networks have shown outstanding functionality in a variety of fields has resulted in their expanding use for decision making in social areas. However, the data required to train a deep neural network is extracted from the real world and, in most cases, inevitably infected with implicit bias. Also, AI systems are developed by decontextualized data, while the meaning of data is correlated to their context. When this context is vague for the AI system, it may carry racial, gender, class, etc. biases and might make mistakes in detecting correlations between things which could lead to biased decisions. In this presentation, we propose an innovative solution based on the Generative Adversarial Network (GAN) to mitigate bias in the outputs of our model. We use two models both of which can be deep neural networks called generator and discriminator. Unlike models such as decision trees, rule-based systems, etc., deep neural networks are not transparent. Whereas some declare that this non-transparency is the problem of DL, in the most interesting cases, impressive scores obtained by DL cannot be retrieved by altering them into transparent models. We argue that detecting bias in outputs of a neural network eventuates to loss of trust, while transparency is just a means to reach the goal of trust. As some experiments show, an increase in transparency may raise a decrease in trust. We suggest a new method to debias the model and attain trust with no need for transparency. In addition to transparency, one needs an explanation about what is going on in the model related to the protected attributes. Our suggested method gives a generic way to protect the interested attributes and hence, provides the user with an explanation.