EthicalML: Injecting Ethical and Legal Constraints into Machine Learning Models

我们选择看什么电影或读什么小说，可能会受到基于机器学习（ML）的推荐系统的建议的影响。然而，在一些重要的场景中，机器学习系统是有缺陷的。下面的每个场景都涉及到这样一种情况：我们希望训练一个机器学习系统，以便它提供服务。然而，在每种情况下，必须对机器学习系统的操作施加一个重要的约束。场景1：我们想要一个系统，将提交的工作申请与我们的学术空缺列表相匹配。该制度必须对少数群体一视同仁。场景2：我们需要一个基于活检图像的自动癌症诊断系统。我们也有艾滋病毒检测结果，可以在培训时使用，但不应该从我们的新患者身上收集。场景3：我们希望有一个系统可以帮助我们决定是否批准抵押贷款申请。我们需要了解决策过程，并将其与我们的清单联系起来，例如申请人在过去三个月内是否有透支，是否在选民名册上。场景1要求机器学习系统做出公平的决定，在种族、性别和残疾等方面不存在歧视；方案2需要机器学习系统来保护个人敏感数据的机密性；场景3要求机器学习系统通过提供人类可以理解的决策来实现透明度。为ML模型配备道德和法律约束（场景1-3）是一个严重的问题；如果没有这一点，机器学习的未来将面临风险。在英国，这一点得到了下议院科学技术委员会的认可，该委员会建议紧急成立数据伦理委员会（“大数据困境”报告，2016年）。此外，自2015年以来，英国皇家学会启动了一个政策项目，研究与ML模型及其用例的进步相关的社会、法律和道德挑战。构建具有公平性、保密性和透明度约束的机器学习模型是一个活跃的研究领域，并且可以使用不相关的框架来解决每个约束。然而，如何把它们放在一起并不明显。我的长期目标是开发一个具有即插即用约束的ML框架，该框架能够处理上述任何约束、它们的组合以及将来可能规定的新约束。提议的ML框架依赖于将道德和法律约束实例化为特权信息。在训练时可以使用这些特权信息，以便更好地训练决策模型并使决策模型具有非歧视性，但在部署时，它将无法用于将来的数据。出于保密限制，个人机密数据，如艾滋病毒检测结果，属于保密信息。对于公平约束，种族和性别等受保护的特征是特权信息。对于透明度约束，复杂的不可解释但高度区分的特征（如深度学习特征）是特权信息。该项目旨在开发一个ML框架，该框架可以产生准确的预测和对其预测的不确定性估计，同时也符合道德和法律约束。这项建议的主要贡献是：1)一种新的特权学习算法，通过在部署时允许即插即用的各种约束，通过核化，通过优化其超参数，以及通过产生预测不确定性的估计，克服了现有方法的局限性；2)一种可扩展和自动推理，使新的特权学习算法易于适用于任何大规模的学习问题，如二元分类，多类分类和回归；3)将公平性、保密性和透明度限制纳入ML模型的新算法的实例化。

Our choice as to which movies to watch or novels to read can be influenced by suggestions made by machine learning (ML)-based recommender systems. However, there are some important scenarios where ML systems are deficient. Each of the following scenarios involves a situation where we wish to train an ML system so that it delivers a service. In each case, however, there is an important constraint that must be imposed on the operation of the ML system.Scenario 1: We want a system that will match submitted job applications to our list of academic vacancies. The system has to be non-discriminatory to minority groups. Scenario 2: We need an automated cancer diagnosis system based on biopsy images. We also have HIV test results, which can be used at training time but should not be collected from our new patients.Scenario 3: We wish to have a system that can aid us in deciding whether or not to approve a mortgage application. We need to understand the decision process and relate it to our checklist such as whether or not the applicant has an overdraft in the last three months and is on electoral roll. Scenario 1 asks an ML system to be fair in its decisions by being non-discriminatory with regards to, e.g., race, gender, and disability; scenario 2 requires an ML system to protect confidentiality of personal sensitive data; and scenario 3 demands transparency from an ML system by providing human-understandable decisions. Equipping ML models with ethical and legal constraints, scenarios 1-3, is a serious issue; without this, the future of ML is at risk. In the UK, this is recognized by the House of Commons Science and Technology Committee, which recommended an urgent formation of a Council of Data Ethics ("The Big Data Dilemma" report, 2016). Furthermore, since 2015, the Royal Society has started a policy project that looks at the social, legal, and ethical challenges associated with advancement in ML models and their use cases.Building ML models with fairness, confidentiality, and transparency constraints is an active research area, and disjoint frameworks are available for addressing each constraint. However, how to put them all together is not obvious. My long-term goal is to develop an ML framework with plug-and-play constraints that is able to handle any of the mentioned constraints, their combinations, and also new constraints that might be stipulated in the future.The proposed ML framework relies on instantiating ethical and legal constraints as privileged information. This privileged information is available at training time to better train a decision model and to make a decision model non-discriminatory, but it will not be accessible for future data at deployment time. For confidentiality constraints, personal confidential data such as HIV test results are the privileged information. For fairness constraints, protected characteristics such as race and gender are the privileged information. For transparency constraints, complex un-interpretable but highly discriminative features such as deep learning features are the privileged information.This project aims to develop an ML framework that produces accurate predictions and uncertainty estimates about its predictions while also complying with ethical and legal constraints. The key contributions of this proposal are: 1) a new privileged learning algorithm that overcomes limitations of existing methods by allowing to plug-and-play various constraints at deployment time, by being kernelized, by optimizing its hyperparameters, and by producing estimates of prediction uncertainty, 2) a scalable and automated inference that makes the new privileged learning algorithm easily applicable for any large scale learning problem such as binary classification, multi-class classification, and regression, and 3) an instantiation of the new algorithm for incorporating fairness, confidentiality, and transparency restrictions into ML models.