随着人工智能（AI）技术的不断进步，越来越多的组织也开始渴望采用这些创新技术。鉴于其不断增长的期望，组织面临的难题是如何将AI技术规模化以实现最佳效果。答案是，组织必须采用并实施MLOps！
MLOps（Machine Learning Operations ），是一种将机器学习模型应用到生产环境中的方法和实践。它涵盖了整个机器学习模型的生命周期，包括模型开发、训练、部署、管理、更新以及监控。MLOps旨在使机器学习模型部署更快、更可靠、更易于管理。同时，MLOps还强调了透明度、可重复性和可维护性，这些都是确保生产环境中的机器学习模型正常运行的重要因素。
MLOps涵盖了一系列与机器学习(ML)部署和管理相关的流程和实践。如果没有足够的MLOps能力，组织可能会难以实现人工智能(AI)的潜力。事实上，许多组织仍然需要赶上其目标AI成熟度，并且由于缺乏专业知识、投资和基础设施，尚未真正实施ML。
MLOps encompasses a set of processes and practises essential to the deployment and management of machine learning (ML). Without sufficient MLOps capabilities, organisations can struggle to realise the potential of AI. Indeed, many organisations still need to catch up to their target AI maturity and have yet to implement ML meaningfully because of a lack of expertise, investment and infrastructure.
为了探讨这个话题，德勤调查了他们组织中了解或负责数据、分析和人工智能的专业人士。这个人工智能成熟的人口统计数据使我们能够衡量领导层和其他知识团体对他们组织内部人工智能状态的看法，以及他们在实施ML时面临的挑战。这个独特的样本揭示了组织如何扩大人工智能解决方案，需要什么，现在正在使用什么技术，以及为未来计划什么技术。
To explore this topic, Deloitte surveyed professionals who are knowledgeable or responsible for data, analytics and AI within their organisations. This AI-mature demographic allowed us to gauge the perception of leadership and other knowledgeable groups on the state of AI within their organisations and the challenges they face in implementing ML. This unique sample reveals how organisations want to scale AI solutions, what is required and what technologies are being used now and planned for the future.
德勤关于通过机器学习运营（MLOps）扩展人工智能的洞察报告，主要内容总结如下——
MLOps的重要性：随着人工智能技术的发展，组织越来越希望利用这些创新技术。为了实现这一目标，组织必须采用和实施MLOps。MLOps是一组旨在可靠、高效地开发、部署和维护生产中的机器学习模型的实践。
AI成熟度：调查显示，大多数组织希望在未来三年内成为行业领先或市场领先，但目前只有15%的组织认为自己的AI成熟度非常高。为了提高AI成熟度，组织必须克服复杂的障碍，并实施适当的MLOps流程和工具。
数据转型和管理：数据转型是ML的关键方面。为了有效部署ML，数据需要以正确的形式存在。组织需要大量的数据，并且收集、管理和存储数据是耗时的。大多数受访者表示，模型开发、数据转型和模型管理和监控是需要最多努力的AI维度。
技术人才需求：MLOps是一个新领域，许多组织缺乏必要的技能集。数据显示，26%的组织缺少MLOps工程师，28%需要更多的IT架构师。
投资和基础设施：尽管云服务在组织中广泛使用，但实施其他技术（如一体化和其他专业MLOps平台）的速度仍然落后。大多数受访者计划在不久的将来进一步投资于广泛的技术。
监管环境：AI和ML的监管环境正在迅速发展，商业领袖需要意识到合规的重要性。MLOps框架可以帮助管理监管环境，同时为消费者、社会和组织提供保障。
投资MLOps的回报：投资MLOps的好处体现在各个方面，从提高员工生产力、提供更好的产品和服务，到减少生产时间和成本，最终实现投资回报（ROI）。实施MLOps的组织实现目标的可能性是其他组织的两倍。
AI成熟度
机构正急切地寻求提升其人工智能（AI）成熟度。据估计，到2025年，AI和机器学习（ML）将为企业带来高达4.4万亿美元的商业价值，而MLOps市场预计将扩大至40亿美元。在我们的样本中，有92%的受访者设定了在未来三年内成为行业或市场领先者的目标。
Organisations are looking to grow their level of AI maturity quickly. By 2025, AI and ML are estimated to drive US$4.4 trillion in business value,2with the MLOps market expected to expand to US$4 billion.3This desire is widespread within our sample, with 92 per cent of respondents setting a target to be industry-leading or market-leading within the next three years.
尽管许多机构都希望在人工智能领域处于行业领先或市场领先地位，但数据显示，这将需要从目前的状态进行重大改进。目前只有15%的组织将其人工智能成熟度评为非常成熟(图1)。
However, while many organisations desire to be industry-leading or market-leading in AI, the data shows that this will require significant improvement from their current state. Only 15 per cent of organisations currently rate their AI maturity as very mature (figure 1).
为了实现人工智能成熟度的提升，企业必须确保自己做好准备，以应对复杂的障碍，并实施适当的MLOps流程和工具。我们还发现，对于企业当前的人工智能成熟度，C级管理层与技术岗位的看法存在差异（见图2）。
For this increase in AI maturity to occur, organisations must ensure they are prepared to negotiate complex barriers and implement adequate MLOps processes and tools. We also see a variation in the perception of an organisation’s current AI maturity between the C-suite and more technical roles (figure 2).
MLOps的重要性
如果组织想要扩大人工智能的应用范围并达到预期的成熟度，就需要克服一些障碍。其中一个就是生产化问题——组织不能再依赖手动流程来实现机器学习。相反，他们需要一种自动化、高效且可扩展的方法。MLOps正是针对这一问题以及日益复杂的机器学习系统和多样性而产生的。对于采用人工智能的组织来说，MLOps是必不可少的，它可以自动化流程和操作，加快机器学习模型的生命周期。
If organisations are to scale AI and reach their target maturity, there are several barriers to overcome. Productionalisation is one issue – organisations can no longer rely on manual processes to bring ML to life.4Instead, they need an automated, efficient and scalable approach. MLOps has emerged as a response to this problem and the growing complexity and diversity of ML systems. It is an essential requirement for organisations utilising AI, automating processes and operations and accelerating the ML model’s life cycle.
MLOps背后的基本原则是一致性，以及AI解决方案构建的可重复方法。MLOps能力创建了端到端的管道和架构，帮助数据科学家、ML工程师和其他开发人员在生产中进行实验和快速部署模型。它创造了软件工程和持续集成和持续交付(CI / CD)原则的文化，以帮助组织将AI解决方案扩展到生产中。MLOps能力定义的健壮模型治理框架确保了生产中模型及其对客户的影响的信心。
The fundamental principle behind MLOps is consistency, and a repeatable approach for AI-solution builds. MLOps capability creates the end-to-end pipeline and architecture that helps data scientists, ML engineers and other developers experiment and rapidly deploy models in production. It creates a culture of software engineering and continuous integration and continuous delivery (CI/CD) principles to help organisations scale AI solutions into production. The robust model governance framework defined by MLOps capability ensures confidence around the models in production and their impact on customers.
AI 领域的发展日新月异。随着新一代 AI 技术的兴起，MLOps 的重要性也随之提升。只有具备高级 MLOps 能力的组织才能利用最热门的 AI 算法。数据显示，组织对探索更复杂的 AI 方法（如强化学习和生成模型）的热情高涨，以挖掘以前未被利用的价值源泉。在受访者中，41%计划明年使用生成模型，42%计划使用强化学习。随着新的技术和算法的出现，组织的潜在应用场景和价值也随之增加。
The AI landscape is fast-moving. As the next generation of AI comes to the fore, the importance of MLOps is heightened. Only organisations with advanced MLOps capability can take advantage of the most sought-after AI algorithms. The data shows organisations’ strong motivation to explore more complex approaches to AI, such as deep-learning techniques like reinforcement learning and generative models, to realise previously untapped sources of value. Of the sample, 41 per cent plan to use generative models next year and 42 per cent plan to use reinforcement learning. As new techniques and algorithms emerge, so do potential use cases and value for organisations.
数据转型与管理
数据转换是ML（机器学习）的一个重要方面。要完全部署ML，数据必须以正确的形式存在。ML需要大量的数据才能有效，而收集、管理和存储数据是一项耗时的工作。事实上，大多数受访者表示，需要投入最多精力的AI维度是模型开发、数据转换和模型管理与监控（图3）。
Data transformation is a critical aspect of ML. To fully deploy ML, data needs to be in the correct form. ML requires large amounts of data to be effective, and it is time-consuming to collect, manage and store. Indeed, the majority of respondents indicated that the AI dimensions requiring the most effort are model development, data transformation and model management and monitoring (figure 3).
AI模型需要高质量的数据，这些数据需要得到良好的管理和适当的转换，但在实践中这可能具有挑战性。通常，像金融服务等已经建立的行业拥有大量以遗留格式存储的数据。这使得它们很难整合并用于现代的ML模型中。
AI models require high-quality data that are well-governed and properly transformed, which can be challenging in practice. Often well-established businesses, like those in the financial services industry, have significant amounts of data stored in legacy formats. This makes it difficult to integrate and use in modern ML models.
例如，企业数据库中的传统数据形式（如数据库、文件和带有非结构化文本的系统）难以利用。一些组织中的用于ML模型的数据复杂且混乱，包含缺失值、离群值和其他异常情况。数据转换涉及清洗和预处理这些数据，这可能需要耗费大量时间。
For example, traditional forms of enterprise data in databases, files and systems with unstructured text are difficult to utilise.5Data used for ML models is complex and messy in some organisations, with missing values, outliers and other anomalies. Data transformation involves cleaning and pre-processing this data, which can be time consuming.
技术人才需求
MLOps是一个新兴领域，许多组织缺乏具备必要技能的人才（如图4所示）。在未来五年内，对技术人才的需求预计将持续存在，其中跨行业的MLOps工程师和IT架构师的需求预计将尤其高。
MLOps is a new field, and many organisations lack staff with the necessary skill sets (figure 4). Demand for technical talent will likely persist in the next five years with demand for MLOps engineers and IT architects across industries expected to be particularly high.
数据显示，26%的组织缺少MLOps工程师， 28%的组织需要更多的IT架构师。这表明，开发MLOps以扩大人工智能能力所需的技术技能存在差距。
Our data shows that 26 per cent of organisations are missing MLOps engineers and 28 per cent need more IT architects. This demonstrates a gap in the technical skill sets required to develop MLOps to scale AI capabilities.
企业必须认识到所需的专业技能。“企业的认同”被认为是组织达到人工智能成熟度的一个障碍。然而，在“认为这是一个障碍”的受访者中，又有85% 的受访者并没有把预算方面的挑战作为障碍。这表明，受访机构面临的认同问题并非是财务问题。这也表明，表现最佳、收入最高的人工智能成熟企业拥有围绕 MLOps 的强大文化，并对其所能带来的成功和价值有共识。
Organisations must recognise the specialist skills required. Business buy-in was identified as an obstacle to reaching an organisation’s target AI-maturity rate. Yet, 85 per cent of survey respondents who saw this as an obstacle did not select budgetary challenges as an obstacle. This suggests the buy-in issues that respondent organisations are facing are predominantly non-financial. It also indicates that the top-performing, highest-revenue AI-mature organisations have a strong culture around MLOps, and a shared understanding of the success and value it can bring.
人工智能正在快速发展，新兴技术将被雄心勃勃的组织利用，以扩大其人工智能能力，并用MLOps迎接挑战。组织应该致力于通过克服数据转换、基础设施落后和投资不足的关键障碍，开发尽可能大的MLOps能力。
AI is developing rapidly, and emerging technologies will be harnessed by ambitious organisations ready to scale their AI capabilities and meet the challenge with MLOps. Organisations should aim to develop the greatest possible MLOps capability by overcoming key barriers of data transformation, lagging infrastructure and lack of investment.
