Algorithmic Fairness in Black-box Machine Learning Models

深度学习模型是成功的机器学习模型，在过去几年中取得了令人难以置信的增长，在许多应用中取得了令人印象深刻的结果。深度学习模型正越来越频繁地用于影响人们生活的领域的决策，如就业，教育，警务和贷款审批。这些用途引起了人们对算法歧视偏见的担忧，并推动了公平意识机器学习的发展。在这个快速发展的领域，最初的努力集中在正规化可用于训练新模型的公平性统计指标上。虽然最初的努力是解决机器学习中公平性问题的重要第一步，但在将它们应用于深度学习模型时，存在直接的挑战。深度学习模型的主要成功来自于高效的学习算法及其具有数百层和数百万参数的巨大参数空间的组合。如此庞大而复杂的参数空间使得深度学习模型被称为黑箱模型。这意味着我们无法看到算法的内部，也无法理解它是如何做出决定的。该研究项目的主要目标是开发数学工具和算法，以实现有效和高效的公平意识深度学习方法。我们的目标是使公平意识方法达到成熟和可靠的水平，任何人都可以信任这些自动化系统，而不必担心歧视。该研究计划将解决以下问题：（a）描述数据如何存在偏差以及这些偏差如何影响深度学习模型。 (b)验证经过训练的深度学习模型的区分度，并说明经过训练的模型如何不公平。(c)通过在学习过程中施加公平约束或消除偏见，在学习中部署公平保证。 (d)在学习和决策中共同执行公平。(e)向最终用户解释最终决定。解释决策是验证过程公平性的必要步骤，可以通过一系列措施来完成，从向最终用户描述决策过程，到提供解决方案以改变不受欢迎的决策。深度学习中算法识别的开发和分析可能是实现未来科学和技术进步的关键。这项研究将有助于加拿大在人工智能领域的领导地位，并推动下游应用，如公平的医疗保健，这可以帮助加拿大人的日常生活。

Deep learning models are successful machine learning models that have seen incredible growth in the last few years with impressive results in many applications. Deep learning models are being used with increasing frequency for decision making in domains that affect peoples' lives such as employment, education, policing and loan approval. These uses raise concerns about biases of algorithmic discrimination and have motivated the development of fairness-aware machine learning. The initial efforts in this fast-growing field were focusing on formalizing statistical measures of fairness that could be used to train new models. While initial efforts were important first steps towards addressing fairness concerns in machine learning, there were immediate challenges when applying them to deep learning models. The main success of deep learning models comes from a combination of efficient learning algorithms and their huge parametric space with hundreds of layers and millions of parameters. Such huge and complex parametric space makes deep learning models to be known as black box models. This means that we are not able to see the inside of an algorithm and understand how it arrives at a decision. The main objective of this research program is to develop mathematical tools and algorithms for effective and efficient fairness-aware deep learning approaches. The goal is to bring fairness-aware approaches to a level of maturity and reliability that anyone can trust these automated systems without a fear of discrimination. This research program will address the followings: (a) Describe how data is biased and how such biases are affecting the deep learning models. (b) Verify discrimination in trained deep learning models and specify how a trained model is not fair. (c) Deploy fairness guarantees in learning by either imposing fairness constraints or removing bias during learning. (d) Enforce fairness jointly in the learning and decision-making. (e) Explain the final decisions to the end users. Explaining the decision is a necessary step to validate the fairness of the process which can be done through a range of measures from describing the process of decision making to the end users, to providing solutions to change the undesirable decisions. The development and analysis of algorithmic discrimination in deep learning can be a key in enabling future scientific and technological progress. This research will contribute to Canada's position as a leader in artificial intelligence and fuel downstream applications, such as fair health care, that can help Canadians in their daily lives.