The age of AI is upon us and many companies begin to start their AI journey and reap the full potential of AI in their respective industries. But, some still consider AI as an immature technology with plenty of ways for it to go wrong. Therefore, before starting your long AI journey, there are some pitfalls you should avoid in implementing and developing AI solutions. They’re a result of the anecdotal, personal and published experience of AI projects that could have gone better.

AI的时代已经过去，许多公司开始了他们的AI之旅，并在各自的行业中收获了AI的全部潜力。 但是，仍然有人认为人工智能是一种不成熟的技术，有很多出错的方法。 因此，在开始漫长的AI之旅之前，在实施和开发AI解决方案时应避免一些陷阱。 它们是人工智能项目的传闻，个人和出版经验的结果，而这些经验本来可以做得更好。

1.建立已经成为行业标准的AI系统 (1. Building AI systems that have become industry standards)

Reinventing the wheel, that’s the reasonable words to describe building an AI system that has become an industry standard. It is a waste of your company’s time and resources. Instead, buy it from a company that has done research and development for years, and has launched a product that has been used and trusted by ample of users. Embrace their solution because this buy decision can get you high-quality AI services at a fraction of the cost and time that it would take to develop in-house. Because building an AI system in-house is a costly and risky endeavor, only do it if the AI system is quite specialized to your business and allow you to build a unique defensible advantage, something that can differentiate your company from its competitors.

重新发明轮子，这是描述构建已成为行业标准的AI系统的合理用语。 这浪费了您公司的时间和资源。 相反，应从经过多年研究和开发的公司购买该产品，并且该产品已被许多用户使用和信任。 拥抱他们的解决方案，因为这个购买决定可以为您提供高质量的AI服务，而所需的成本和时间仅为内部开发的一小部分。 由于内部构建AI系统是一项昂贵且冒险的工作，因此只有在AI系统非常适合您的业务并允许您建立独特的防御优势时才可以这样做，这可以使您的公司与竞争对手区分开来。

2.使用人工智能使工作自动化 (2. Using AI to automate jobs)

Most existing AI systems have a narrow capability, i.e. it is programmed to perform a single task — whether it’s detecting the position of vehicles on the road, recognizing faces, predicting the weather, or analyzing the movement of NBA players. As a result, these systems don’t perform outside of the single task that they are designed to perform. Thus, you shouldn’t expect a single AI system to automate a job.

现有的大多数AI系统的能力都很有限，即被编程为执行一项任务-无论是检测车辆在道路上的位置，识别人脸，预测天气还是分析NBA球员的运动。 结果，这些系统不会在其设计要执行的单个任务之外执行。 因此，您不应该期望单个AI系统自动完成一项工作。

A job normally consists of several tasks that are associated with each other. What you can do, at the beginning, is to select a single task to be performed by an AI system. Find a task that is the key driver, the main pain point, or the bottleneck of the business workflow.

作业通常由相互关联的多个任务组成。 首先，您可以做的是选择一个要由AI系统执行的任务。 查找成为关键因素，主要痛点或业务工作流程瓶颈的任务。

Despite its narrow capability, we shouldn’t underestimate it. AI systems can improve our work productivity and efficiency due to their ability to process data and complete tasks at a significantly quicker pace than any human being can. Consequently, the bottleneck can be reduced, the workflow can be continuing smoothly, and boring, routine, and mundane tasks can be downsized.

尽管功能有限，但我们不应该低估它。 人工智能系统可以以比任何人更快的速度处理数据和完成任务的能力，从而提高了我们的工作效率和效率。 因此，可以减少瓶颈，工作流程可以顺利进行，并且可以减少无聊的，日常的和平凡的任务。

3.数据不足和数据质量差 (3. Lack of data and poor data quality)

Most of AI systems use deep learning as their foundational technology. Its main limitation: it requires thousands, millions, or even billions of training examples in order to perform a particular task. If the AI system that you need is an application that relies on supervised learning, you need to ask, at the very least, the following questions:

大多数AI系统都将深度学习作为其基础技术。 它的主要局限性是：它需要成千上万，数百万甚至数十亿的训练示例才能执行特定任务。 如果您需要的AI系统是依赖监督学习的应用程序，那么您至少需要问以下问题：

1. Do you have the data to train the AI system?您是否拥有训练AI系统的数据？
2. Are the quality and the amount of data that you have adequate to achieve the expected performance from the system?您是否有足够的质量和数据量来实现系统的预期性能？
3. If the existing data is inadequate qualitatively and/or quantitatively, do you have the capacity and capability to acquire more data?如果现有数据在质量和/或数量上不足，您是否有能力和能力来获取更多数据？

If the answer to one or more of the above questions is no, you should rethink about your own AI initiative. The quality and amount of data oftentimes determines the performance of deep learning systems. Thus, when you don’t have enough quality training data, deep learning would fail miserably and might not be your ideal solution.

如果对以上一个或多个问题的回答为否，则应重新考虑自己的AI计划。 数据的质量和数量通常决定着深度学习系统的性能。 因此，当您没有足够质量的培训数据时，深度学习将惨遭失败，并且可能不是您的理想解决方案。

4.对人工智能的期望不切实际 (4. Having unrealistic AI expectations)

I blame sci-fi movies for causing people to have unreasonable expectations towards AI. AI has indeed surpassed human in performing particular tasks, making the impression that AI will take over the world soon. But, to think AI can surpass human in any tasks is irrational. Similarly, to think AI can cure any problems encountered by your company is unrealistic. So, it’s important to know what’s currently feasible and what isn’t.

我指责科幻电影导致人们对AI抱有不合理的期望。 在执行特定任务时，人工智能确实已经超越了人类，给人的印象是，人工智能将很快占领整个世界。 但是，认为AI在任何任务中都能超越人类是不合理的。 同样，认为AI可以解决您公司遇到的任何问题也是不现实的。 因此，了解当前可行和不可行很重要。

“Anything you can do with one second of thought, we can probably now or soon automate”

“您只要花一秒钟的时间就能完成任何事情，我们现在可能会或即将实现自动化”

- Andrew Ng

-吴安达

AI tends to work well when the task is simple and there is a lot of data available. To give you a rule of thumb, anything you can do with one second of thought, we can probably now or soon automate with AI. So, it’s still a long way until AI can write you 80 pages market research report. The reason is because AI needs tonnes of data and loads of tries to succeed on very specific problems, and it is difficult to generalize its knowledge on tasks very different to those trained upon. AI can learn, but it won’t suddenly learn all aspects of human intelligence and outsmart us.

当任务很简单并且有大量可用数据时，AI往往会运行良好。 为了给您一个经验法则，您只要花一秒钟的时间就可以完成任何事情，我们现在可能或即将使AI自动化。 因此，距离AI可以为您撰写80页的市场研究报告还有很长的路要走。 原因是因为AI需要大量的数据和大量尝试才能成功解决非常具体的问题，并且很难将其知识推广到与受过培训的任务截然不同的任务上。 人工智能可以学习，但不会突然学习人类智能的所有方面并超越我们。

Knowing what the limitation of AI is crucial for succesful implementation of AI. Don’t expect AI to solve everything and be realistic about what AI can and can’t do

知道AI的局限性对于成功实施AI至关重要。 不要指望AI解决所有问题并对AI可以做什么和不能做什么具有现实性

5.未能使AI项目与业务目标保持一致 (5. Failing to align AI projects with business goals)

Don’t make AI implementation as the primary goal as it can derail your business from its original goals. Thus, start an AI initiative by combining AI knowledge and domain knowledge, i.e. select projects that can be done with AI and valuable for your business.

不要将AI实施作为主要目标，因为它会使您的业务偏离其最初目标。 因此，通过结合AI知识和领域知识来启动AI计划，即选择可以用AI完成并对您的业务有价值的项目。

It’s like combining AI knowledge and domain knowledge and finding a sweet spot between those two. AI knowledge is mainly possessed by your AI and machine learning engineers and domain knowledge by your business associates. Therefore, don’t count solely on machine learning engineers to come up with the use cases of AI. Instead, pair engineering talent with business talent and work cross-functionally to find feasible and valuable projects.

这就像将AI知识和领域知识相结合，并在两者之间找到了一个完美的结合点。 AI知识主要由AI和机器学习工程师拥有，而业务伙伴则具有领域知识。 因此，不要仅仅依靠机器学习工程师来提出AI的用例。 相反，应将工程人才与业务人才配对，并跨职能开展工作，以找到可行且有价值的项目。

6.期望AI计划在第一次尝试时就起作用 (6. Expecting AI initiative to work at the first try)

Machine learning algorithms, particularly neural networks, are often thought of as black boxes due to the convoluted nature of the processes between their input and output. Input data undergo complex transformations in multiple layers of the algorithm, which cause the model to behave in complex and unpredictable ways. If our model doesn’t work as expected, we can’t manually tuning its parameters to fix the problem. What we can do is by feeding the model with more fine-quality data and/or tune the model hyperparameters.

机器学习算法，特别是神经网络，由于其输入和输出之间过程的复杂性，通常被认为是黑匣子。 输入数据在算法的多个层中经历复杂的转换，这导致模型以复杂且不可预测的方式运行。 如果我们的模型无法按预期运行，则无法手动调整其参数来解决问题。 我们可以做的是为模型提供更多高质量的数据和/或调整模型的超参数。

Uncertainties are inherent in an AI project. Thus, don’t expect an AI project to work the first time. Instead, plan AI development to be an iterative process, with multiple attempts needed to succeed.

不确定性在AI项目中是固有的。 因此，不要指望一个AI项目会在第一时间工作。 相反，将AI开发计划为迭代过程，需要多次尝试才能成功。

Due to its infancy, the majority of AI projects fail. Knowing the pitfalls of AI implementation is one of the important first steps in your company’s AI journey. Attention to the pitfalls will help us in solving any problems along the way and achieve a succesful AI initiative.

由于其起步阶段，大多数AI项目都失败了。 了解AI实施的陷阱是您公司AI旅程中重要的第一步。 注意陷阱将帮助我们解决整个过程中的任何问题并实现成功的AI计划。