您现在访问的是微软AZURE全球版技术文档网站,若需要访问由世纪互联运营的MICROSOFT AZURE中国区技术文档网站,请访问 https://docs.azure.cn.

负责任且受信任的 AIResponsible and trusted AI

Microsoft 概述了六个关键原则,其中包含责任 AI:责任、inclusiveness、可靠性和安全性、公平、透明性以及隐私和安全性。Microsoft outlines six key principles for responsible AI: accountability, inclusiveness, reliability and safety, fairness, transparency, and privacy and security. 在将受信任的 AI 转移到更多主流产品和服务时,这些原则至关重要。These principles are essential to creating responsible and trustworthy AI as it moves into more mainstream products and services. 它们通过两种方式进行指导:道德和 explainable。They are guided by two perspectives: ethical and explainable.

负责 AI 原则的关系图。

规范Ethical

从道德角度来看,AI 应该是公平的并且包含在其断言中,负责其决策,并且不区分或阻碍不同的比赛、残障或背景。From an ethical perspective, AI should be fair and inclusive in its assertions, be accountable for its decisions, and not discriminate or hinder different races, disabilities, or backgrounds.

Microsoft 在2017中建立了工程和研究 Aether中的 AI、道德和效果的道德委员会。Microsoft established an ethical committee for AI, ethics, and effects in engineering and research Aether, in 2017. 这一委员会的核心责任是向负责的问题、技术、流程和最佳实践提出建议。The core responsibility of the committee is to advise on responsible issues, technologies, processes, and best practices. 在此 Microsoft Learn 模块上了解有关 Aether 的详细信息。Learn more about Aether at this Microsoft Learn module.

问责制Accountability

责任是责任 AI 的基本支柱。Accountability is an essential pillar of responsible AI. 设计和部署 AI 系统的人员需要对其操作和决策有责任,尤其是我们对更自治系统的发展进度。The people who design and deploy the AI system need to be accountable for its actions and decisions, especially as we progress toward more autonomous systems. 组织应考虑建立内部审查正文,提供有关开发和部署 AI 系统的监管、见解和指导。Organizations should consider establishing an internal review body that provides oversight, insights, and guidance about developing and deploying AI systems. 尽管本指南可能因公司和区域而异,但它应该反映出组织的 AI 旅程。While this guidance might vary depending on the company and region, it should reflect an organization's AI journey.

包容Inclusiveness

Inclusiveness 要求 AI 应考虑所有人为争用和体验,而非包含设计实践可帮助开发人员了解并解决可能会无意中排除人员的潜在障碍。Inclusiveness mandates that AI should consider all human races and experiences, and inclusive design practices can help developers to understand and address potential barriers that could unintentionally exclude people. 在可能的情况下,应使用语音转文本、文本到语音转换和视觉识别技术来为人们提供听觉、视觉和其他障碍。Where possible, speech-to-text, text-to-speech, and visual recognition technology should be used to empower people with hearing, visual, and other impairments.

可靠和安全。Reliability and safety

AI 系统需要可靠和安全才能获得信任。AI systems need to be reliable and safe in order to be trusted. 这对于系统的执行是非常重要的,因为它是最初设计的,并且是为了安全应对新的情况。It's important for a system to perform as it was originally designed and for it to respond safely to new situations. 其固有的复原能力应该不会对预期的或意外的操作产生抵触Its inherent resilience should resist intended or unintended manipulation. 应为操作条件建立严格的测试和验证,以确保系统安全地响应边缘情况,并将 A/B 测试和冠军/争取冠军宝座方法集成到评估过程中。Rigorous testing and validation should be established for operating conditions to ensure that the system responds safely to edge cases, and A/B testing and champion/challenger methods should be integrated into the evaluation process.

AI 系统的性能可能会随着时间的推移而降低,因此,需要在被动中建立可靠的监视和模型跟踪过程,并主动衡量模型的性能,并根据需要重新训练它以实现现代化。An AI system's performance can degrade over time, so a robust monitoring and model tracking process needs to be established to reactively and proactively measure the model's performance and retrain it, as necessary, to modernize it.

什么是 explainableWhat is explainable

Explainability 可帮助数据科学家、审计员和业务决策者确保 AI 系统能够合理地论证他们的决策及其得出的结论。Explainability helps data scientists, auditors, and business decision makers to ensure that AI systems can reasonably justify their decisions and how they reach their conclusions. 这还可以确保遵守公司政策、行业标准和政府法规。This also ensures compliance with company policies, industry standards, and government regulations. 数据科研人员应能够向利益干系人说明他们如何实现某些级别的准确性以及影响此结果的因素。A data scientist should be able to explain to the stakeholder how they achieved certain levels of accuracy and what influenced this outcome. 同样,若要遵守公司的政策,审核员需要使用一种工具来验证模型,而业务决策者需要能够提供透明模型才能获得信任。Likewise, in order to comply with the company's policies, an auditor needs a tool that validates the model, and a business decision maker needs to be able to provide a transparent model in order to gain trust.

Explainability 工具Explainability tools

Microsoft 开发了 InterpretML,这是一个开源工具包,可帮助实现模型 explainability 并支持玻璃箱和黑白模型。Microsoft has developed InterpretML, an open-source toolkit that helps to achieve model explainability and supports glass-box and black-box models.

  • 玻璃框模型是可解释的,因为它们的结构。Glass-box models are interpretable because of their structure. 对于这些模型,请使用 Explainable 提升计算机,它是基于决策树或线性模型的算法的状态,提供无损说明,可由域专家编辑。For these models, use Explainable Boosting Machine, which is the state of the algorithm based on a decision tree or linear models, provides lossless explanations, and is editable by domain experts.

  • 由于复杂的内部结构(神经网络),因此很难解释黑色框模型。Black-box models are more challenging to interpret because of a complex internal structure, the neural network. Explainers 如酸橙色或 SHapley 加法说明 (SHAP) 通过分析输入和输出之间的关系来解释这些模型。Explainers like LIME or SHapley Additive exPlanations (SHAP) interpret these models by analyzing the relationship between the input and output.

  • Fairlearn 是 SDK 和 AutoML 图形用户界面的 Azure 机器学习集成和开源工具包。Fairlearn is an Azure Machine Learning integration and an open-source toolkit for the SDK and the AutoML graphical user interface. 使用 explainers 了解主要影响模型和域专家验证这些影响的因素。Use explainers to understand what mainly influences the model and domain experts to validate these influences.

在 Azure 机器学习中浏览模型 interpretability ,详细了解 explainability。Explore model interpretability in Azure Machine Learning to learn more about explainability.

公平Fairness

公平是一种核心道德原则,所有人都以此来理解和应用。Fairness is a core ethical principle that all humans aim to understand and apply. 开发 AI 系统时,此原则更为重要。This principle is even more important when AI systems are being developed. 关键检查和余额需要确保系统的决策不会对一个组或个人的性别、种族、性取向或宗教偏差进行区分或运行。Key checks and balances need to make sure that the system's decisions don't discriminate or run a gender, race, sexual orientation, or religion bias toward a group or individual.

  • Microsoft 提供 ai 公平清单 ,为 ai 系统提供指导和解决方案。Microsoft provides an AI fairness checklist that offers guidance and solutions for AI systems. 这些解决方案分为以下几个阶段:构想、原型、构建、启动和发展。These solutions are loosely categorized into five stages: envision, prototype, build, launch, and evolve. 每个阶段都列出了建议的截止项活动,这些活动有助于最大程度地减少系统中 unfairness 的影响。Each stage lists recommended due diligence activities that help to minimize the impact of unfairness in the system.

  • Fairlearn 与 Azure 机器学习集成,并支持数据科学家和开发人员评估和提高其 AI 系统的公平。Fairlearn integrates with Azure Machine Learning and supports data scientists and developers to assess and improve the fairness of their AI systems. 工具箱提供各种 unfairness 缓解算法,以及直观显示模型公平的交互式仪表板。The toolbox provides various unfairness mitigation algorithms and an interactive dashboard that visualizes the fairness of the model. 在构建模型时,使用工具包并密切评估模型的公平;这应该是数据科学过程中不可或缺的一部分。Use the toolkit and closely assess the fairness of the model while it's being built; this should be an integral part of the data science process.

了解如何 在机器学习模型中缓解公平Learn how to mitigate fairness in machine learning models.

透明Transparency

实现透明度有助于团队了解用于定型模型的数据和算法、应用了哪些转换逻辑、最终生成的模型及其关联的资产。Achieving transparency helps the team to understand the data and algorithms used to train the model, what transformation logic was applied to the data, the final model generated, and its associated assets. 此信息提供了有关如何创建模型的见解,从而使其能够以透明方式重现。This information offers insights about how the model was created, which allows it to be reproduced in a transparent way. Azure 机器学习工作区中的快照通过记录或重新训练试验中涉及的所有培训相关资产和指标来支持透明度。Snapshots within Azure Machine Learning workspaces support transparency by recording or retraining all training-related assets and metrics involved in the experiment.

隐私和安全性Privacy and security

数据持有者负责保护 AI 系统中的数据,并且隐私和安全性是此系统不可或缺的一部分。A data holder is obligated to protect the data in an AI system, and privacy and security are an integral part of this system. 个人需要安全,应以不会损害个人隐私的方式进行访问。Personal needs to be secured, and it should be accessed in a way that doesn't compromise an individual's privacy. Azure 差分隐私 通过随机化进程数据并添加噪音来保护和保留隐私,以隐藏数据科学家的个人信息。Azure differential privacy protects and preserves privacy by randomizing data and adding noise to conceal personal information from data scientists.

人类 AI 指导原则Human AI guidelines

人类 AI 设计准则包含18个时间段内发生的18个原则:最初、发生错误时和一段时间。Human AI design guidelines consist of 18 principles that occur over four periods: initially, during interaction, when wrong, and over time. 这些原则旨在产生更具包容性和以人为中心的 AI 系统。These principles are designed to produce a more inclusive and human-centric AI system.

起始阶段Initially

  • 阐明系统可以执行的操作。Clarify what the system can do. 如果 AI 系统使用或生成指标,则必须将其全部显示和跟踪。If the AI system uses or generates metrics, it's important to show them all and how they're tracked.

  • 阐明系统可以执行的操作。Clarify how well the system can do what it can do. 帮助用户了解 AI 不会完全准确,并为 AI 系统可能导致的错误设置预期。Help users to understand that AI will not be completely accurate, and set expectations for when the AI system might make mistakes.

交互期间During interaction

  • 显示根据上下文相关信息。Show contextually relevant information. 提供与用户当前上下文和环境相关的视觉信息,如附近的酒店,并返回接近目标目标和日期的详细信息。Provide visual information related to the user's current context and environment, such as nearby hotels and return details close to the target destination and date.

  • 缓解社会偏差。Mitigate social biases. 请确保语言和行为不会引入意外的构造型或偏置。Make sure that the language and behavior don't introduce unintended stereotypes or biases. 例如,自动完成功能需要同时确认这两个性别。For example, an autocomplete feature needs to acknowledge both genders.

如果错误When wrong

  • 支持高效消除。Support efficient dismissal. 提供一种简单的机制来忽略或消除不良功能/服务。Provide an easy mechanism to ignore or dismiss undesirable features/services.
  • 支持有效更正。Support efficient correction. 提供了一种直观的方式,使其更易于编辑、优化或恢复。Provide an intuitive way of making it easier to edit, refine, or recover.
  • 清楚地说明系统执行该操作的原因。Make clear why the system did what it did. 优化 explainable AI,为 AI 系统的断言提供见解。Optimize explainable AI to offer insights about the AI system's assertions.

随时间推移Over time

  • 记住最近的交互。Remember recent interactions. 保留交互历史记录,以供将来参考。Retain a history of interactions for future reference.
  • 了解用户的行为。Learn from user behavior. 基于用户的行为个性化交互。Personalize the interaction based on the user's behavior.
  • 慎重更新并改编。Update and adapt cautiously. 限制中断性更改,并基于用户的配置文件进行更新。Limit disruptive changes, and update based on the user's profile.
  • 鼓励细化反馈。Encourage granular feedback. 收集用户与 AI 系统交互的反馈。Gather user feedback from their interactions with the AI system.

以角色为中心的可信 AI 框架A persona-centric, trusted AI framework

以角色为中心的可信 AI 框架的关系图。

AI 设计器AI designer

AI 设计器生成模型,并负责:The AI designer builds the model and is responsible for:

  • 数据偏移和质量检查。Data drift and quality checks. 它们检测离群值并执行数据质量检查来确定缺失值、标准化分布、检查数据以及生成用例和项目报告。They detect outliers and perform data quality checks to identify missing values, standardize distribution, scrutinize data, and produce use case and project reports.

  • 评估系统源中的数据,以识别潜在的偏差。Assessing data in the system's source to identify potential bias.

  • 设计 AI 算法来最大程度地减少数据偏移,如发现装箱、分组和规范化 (尤其是在传统的机器学习模型中,例如基于树的) ,可以消除数据中的少数组。Designing AI algorithms to minimize data biases, such as discovering how binning, grouping, and normalization (especially in traditional machine learning models like tree-based ones) can eliminate minority groups from data. 分类 AI 设计重述数据偏移:将) 的行业行业中的社会、种族歧视和性别类分组,这些类依赖于受保护的健康信息 (PHI 以及 (PII) 的个人身份信息。Categorical AI design reiterates data biases by grouping social, racial, and gender classes in industry verticals that rely on protected health information (PHI) and personally identifiable information (PII).

  • 优化监视和警报来识别目标泄漏并增强模型的开发。Optimizing monitoring and alerts to identify target leakage and strengthen the model's development.

  • 确定报表和深入了解模型的最佳实践,提供对模型的精细了解,并避免使用功能或矢量重要性、UMAP 聚类分析、Friedman 的 H 统计、功能效果等的黑白方法。Establishing best practices for reporting and insights that offer a granular understanding of the model and avoiding black-box approaches that use feature or vector importance, UMAP clustering, Friedman's H-statistic, feature effects, and others. 标识指标有助于定义复杂和新式数据集中的关联之间的预测影响、关系和依赖关系。Identification metrics help to define predictive influence, relationships, and dependencies between correlations in complex and modern datasets.

AI 管理员和官员AI administrators and officers

AI 管理员和官员监察 AI、监管和审核框架操作和性能指标,增加了 AI 安全性的实现方式以及业务的投资回报。The AI administrator and officers oversee AI, governance, and audit framework operations and performance metrics, plus how AI security is implemented and the business' return on investment.

  • 监视有助于模型监视的跟踪面板,为生产模型组合模型度量值,并侧重于准确性、模型降级、数据偏移量、偏差以及在推断速度/错误时的变化。Monitoring a tracking dashboard that assists model monitoring, combines model metrics for production models, and focuses on accuracy, model degradation, data drift, deviation, and changes in speed/error of inference.

  • 实现灵活的部署和重新部署 (最好的 REST API) 允许将模型实现为开放式、不可知的体系结构,该体系结构将模型与业务流程相集成,并为反馈循环生成值。Implementing flexible deployment and redeployment (preferably, REST API) allows models to be implemented into open, agnostic architecture, which integrates the model with business processes and generates value for feedback loops.

  • 努力构建模型管理和访问集边界,降低业务和运营的负面影响。Working toward building model governance and access sets boundaries and mitigates negative business and operational impact. 基于角色的访问控制标准确定安全控制,这些控制可保留受限的生产环境和 IP。Role-based access control standards determine security controls, which preserve restricted production environments and the IP.

  • 使用 AI 审核和符合性框架跟踪模型的开发和更改,以满足行业特定的标准。Using AI audit and compliance frameworks to track how models develop and change to uphold industry-specific standards. 可解释和负责的 AI 建立在 explainability 度量、简洁功能、模型可视化效果和行业垂直语言之上。Interpretable and responsible AI is founded on explainability measures, concise features, model visualizations, and industry-vertical language.

AI 业务使用者AI business consumer

AI 业务消费者 (业务专家) 关闭反馈循环,并为 AI 设计器提供输入。AI business consumers (business experts) close the feedback loop and provide input for the AI designer. 预测性决策和潜在偏差,如公平和道德的措施、隐私和合规性,以及业务效率有助于评估 AI 系统。Predictive decision-making and potential bias implications like fairness and ethical measures, privacy and compliance, and business efficiency help to evaluate AI systems.

  • 反馈循环属于业务生态系统。Feedback loops belong to a business' ecosystem. 显示模型偏置、错误、预测速度和公平的数据在 AI 设计器、管理员和监察器之间建立信任和平衡。Data showing a model's bias, errors, prediction speed, and fairness establish trust and balance between the AI designer, administrator, and officers. 以人为中心的评估应逐步提高 AI,并最大程度地减少从多维、复杂数据 (的 AI 学习) 可帮助防止有偏差的学习。Human-centric assessment should gradually improve AI over time, and minimizing AI learning from multidimensional, complex data (LO-shot learning) can help to prevent biased learning.

  • 使用 interpretability 设计和工具,使 AI 系统负责潜在的偏差。Using interpretability design and tools hold AI systems accountable for potential biases. 应该标记模型偏向和公平问题,并将其送到警报和异常检测系统,该系统可从该行为中获知并自动解决偏移。Model bias and fairness issues should be flagged and fed to an alerting and anomaly detection system that learns from this behavior and automatically addresses biases.

  • 每个预测值都应按重要性或影响细分为各个特征或矢量,并提供可导出到业务报表中以进行审核和符合性审查、客户透明度和业务准备的全面预测说明。Each predictive value should be broken down into individual features or vectors by importance or impact and deliver thorough prediction explanations that can be exported into a business report for audit and compliance reviews, customer transparency, and business readiness.

  • 由于增加了全球安全和隐私风险,在推断过程中解决数据冲突的最佳实践需要遵从单个行业行业中的法规;例如,有关不符合 PHI 和 PII 的警报,违反了国家安全法律等。Due to increasing global security and privacy risks, best practices for resolving data violations during inference require complying with regulations in individual industry verticals; for example, alerts about noncompliance with PHI and PII, violation of national security laws, and more.

后续步骤Next steps

探索 人类 ai 准则 ,详细了解责任 ai。Explore human AI guidelines to learn more about responsible AI.