FAI: Towards Fairness in Deep Neural Networks with Learning Interpretation

深度神经网络（DNN）在计算机视觉和自然语言处理等广泛应用中取得了巨大成功。不幸的是，DNN中广泛存在对少数群体的固有歧视。为了促进深度学习的公平性，该项目旨在解决设计，评估和部署DNN系统中算法歧视的挑战性问题。该项目的成功结果将导致在提供理论理解和实用算法方面取得进展，以实现复杂深度学习模型和预测的公平性。该教育计划整合了机器学习，工业统计和社会科学，旨在培养学生掌握信息系统中的数据分析技术，吸引代表性不足的群体成员从事STEM职业。该项目的主要目标是通过利用机器学习生命周期中关键元素的可解释性，系统地研究并促进深度神经网络的公平性，包括建模，数据准备和特征工程。具体而言，所提出的框架从以下几个方面揭示了深度学习中公平性的内在属性。辅助训练目标被设计为正则化增强的局部解释，以促进经典DNN架构的公平性。数据构建和数据增强方法是为了重建DNN训练的公平数据集而开发的。通过识别应用中的敏感特征，提取领域知识，并进一步发展强化学习，以优化现实约束下的模型公平性。最后，建议的研究创新可以嵌入到基于DNN的真实的系统中，例如医疗诊断和推荐系统，并提供具体的解决方案和评估措施。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Deep neural networks (DNNs) have achieved great successes in a wide range of applications such as computer vision and natural language processing. Unfortunately, inherent discrimination widely exists in DNNs towards minority subgroups. To facilitate fairness in deep learning, this project is to tackle the challenging problem of algorithmic discrimination in designing, evaluating, as well as deploying DNN systems. The successful outcome of this project will lead to advances in providing theoretical understandings and practical algorithms to enable fairness in complicated deep learning models and predictions. The education program that integrates machine learning, industrial statistics, and social sciences is to train students with data analytics technologies in information systems, to attract members of underrepresented groups to pursue careers in STEM.The primary goal of this project is to systematically investigate and facilitate fairness in deep neural networks by leveraging the interpretability of key elements in a machine learning life-cycle including modeling, data preparation and feature engineering. Specifically, the proposed frameworks uncover the intrinsic properties of fairness in deep learning from the following aspects. Auxiliary training objectives are designed to regularize the augmented local interpretation to promote the fairness of classical DNN architectures. Data construction and data augmentation approaches are developed towards reconstructing a fair dataset for DNN training. Through identifying sensitive features in applications, domain knowledge is extracted and reinforcement learning is further developed to optimize the model fairness under realistic constraints. Finally, the proposed research innovations could be embedded in DNN based real systems, such as medical diagnosis and recommender systems, with concrete solutions and evaluation measurements.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.