Diverse Adversaries for Mitigating Bias in Training
This work addresses bias mitigation in language models, offering a more stable and effective solution for fairness in AI applications, though it appears incremental in advancing adversarial techniques.
The paper tackled the problem of mitigating bias in language models by proposing a novel adversarial learning approach using multiple diverse discriminators, which substantially improved bias reduction and training stability over standard methods.
Adversarial learning can learn fairer and less biased models of language than standard methods. However, current adversarial techniques only partially mitigate model bias, added to which their training procedures are often unstable. In this paper, we propose a novel approach to adversarial learning based on the use of multiple diverse discriminators, whereby discriminators are encouraged to learn orthogonal hidden representations from one another. Experimental results show that our method substantially improves over standard adversarial removal methods, in terms of reducing bias and the stability of training.