Testing and debugging DNNs for fairness and discrimination

What is it about?

This paper introduces DICE, An information-theoretic tool to test and debug discrimination in DNNs. The key goal of DICE is to assist software developers in triaging fairness defects by ordering them by their severity. Towards this goal, we quantify fairness regarding protected information (in bits) used in decision-making. A quantitative view of fairness defects not only helps in ordering these defects, our empirical evaluation shows that it improves the search efficiency due to the resulting smoothness of the search space. Guided by the quantitative fairness, we present a causal debugging framework to localize inadequately trained layers and neurons responsible for fairness defects.

Featured Image

Why is it important?

Our approach utilizes information theory to quantify the severity of bias. It provides a quantitative individual discrimination.