专家库管理软件专家评审系统_专家系统2 0

专家库管理软件专家评审系统

Are neural networks and deep learning the keys that will finally unlock the path to artificial general intelligence, or are we on the verge of another AI Winter?

是神经网络和深度学习的钥匙最终将打开通往人工智能的道路，还是我们正处于另一个AI Winter的边缘？

If you’re an entrepreneur, then the answer is: it doesn’t matter. Your funding doesn’t depend on grant providers being excited about your research agenda producing a talking robot. It depends on providing real value to real people.

如果您是企业家，那么答案是： 没关系 。 您的资金并不取决于资助提供者对您的研究议程产生一个会说话的机器人而感到兴奋。 这取决于为真实的人提供真实的价值。

The truth is, even though developments in AI so far have yet to produce machines that are intelligent in a way comparable to our own intelligence, the technologies that have been developed along the way are all incredibly useful and full of potential value. And while it seems like neural networks alone aren’t enough to build the most intelligent systems, their full potential to enrich our lives has yet to be fully tapped into.

事实是，尽管迄今为止AI的发展还没有以与我们自己的智能相当的方式生产出智能的机器，但沿途开发的技术都非常有用且具有潜在价值。 尽管仅靠神经网络似乎还不足以构建最智能的系统，但其丰富我们的生活的全部潜力尚未得到充分挖掘。

Today, I’d like to remind us all of a useful and relevant technology from a previous era of AI: expert systems. Though no longer a glamorous research poster child, this technology has made an enduring impact on society. Neural networks are poised to do the same, even without the reinforcement learning, training curricula, and other innovations taking place at the bleeding edge.

今天，我想提醒我们大家一个过去AI时代的有用和相关技术： 专家系统 。 尽管不再是迷人的研究先驱，但这项技术已对社会产生了持久的影响。 即使没有强化学习，培训课程和其他创新手段，神经网络也可以做到这一点。

As we’ll see, just as expert systems automate inferences based on descriptive knowledge, neural networks can do the same for procedural knowledge. And unlike commercial expert systems, which largely remain in the hands of corporations, we stand ready to make neural network systems available to all.

正如我们将看到的，就像专家系统基于描述性知识自动进行推理一样，神经网络也可以对过程知识进行相同的处理 。 与大多数企业掌握的商业专家系统不同，我们随时准备将神经网络系统提供给所有人。

专家系统：回顾 (Expert Systems: A Review)

Expert systems require at least two components to function:

专家系统至少需要两个组件才能运行：

1. A knowledge base
   知识库
2. An inference engine
   推理引擎

The knowledge base stores propositions — which are encoded pieces of information — while the inference engine uses rules of inference to derive new propositions.

知识库存储命题(即信息的编码片段)，而推理引擎则使用推理规则来推导新的命题。

Obviously, a main challenge in building expert systems is the knowledge acquisition needed to build the knowledge base. Domain experts are often in high demand, and themselves usually lack the ability to codify their own knowledge. This means that constructing a knowledge base is a time-consuming collaboration between programmers and experts.

显然，构建专家系统的主要挑战是构建知识库所需的知识获取。 领域专家经常有很高的需求，他们通常缺乏整理自己的知识的能力。 这意味着构建知识库是程序员和专家之间耗时的协作。

However, this scarcity in expertise is also what makes expert systems valuable. Once encoded, a knowledge base is easily transferred, and does not expire. It can be copied endlessly and used across many applications. Expert systems automate difficult technical inferences and knowledge lookups in return for a front-loaded investment of time and effort.

但是，专业知识的匮乏也使专家系统变得有价值。 编码后，知识库很容易转移，并且不会过期。 可以无限复制它，并在许多应用程序中使用它。 专家系统可以自动进行困难的技术推论和知识查找，以换取前期投入的时间和精力。

It should come as no surprise that expert systems were typically found in domains such as medical science, law, and other areas requiring specialized yet explicit knowledge of facts and procedures.

毫不奇怪，专家系统通常出现在诸如医学，法律和其他领域，这些领域需要对事实和程序具有专门而明确的知识。

Since they are a mature technology, rather than learn the technical details it will serve you much better to have an overall understanding of their purpose, requirements, and relevance. We’ll come back to these three factors when discussing neural networks, so keep them in mind.

由于它们是一项成熟的技术，因此与其了解技术细节，不如对它们的用途 ， 要求和相关性有一个整体的了解，这将为您提供更好的服务。 在讨论神经网络时，我们将回到这三个因素，因此请牢记它们。

We’ve already covered the purpose and requirements…

我们已经介绍了目的和要求...

Purpose: expert systems automate reasoning with explicit facts and rules of inference.

目的：专家系统利用明确的事实和推理规则自动进行推理。

Requirements: to build or update expert systems requires the time and knowledge of domain experts, and the ability to represent that knowledge in code.

要求：建立或更新专家系统需要领域专家的时间和知识，以及在代码中表示该知识的能力。

So, are expert systems still relevant? Certainly.

那么，专家系统仍然有意义吗？ 当然。

While it is less common to find pieces of software explicitly called “expert systems” these days, the ideas and implementations of knowledge representation and inference are still very much in use. For example, ROSS, a legal system powered by IBM Watson combines an enormous knowledge base with more modern ways of filtering for relevant information from the field of natural language processing.

尽管如今很少找到明确称为“专家系统”的软件，但是知识表示和推理的思想和实现仍在使用中。 例如，由IBM Watson提供支持的法律系统ROSS结合了庞大的知识库和更现代的方式来过滤自然语言处理领域中的相关信息。

Also, many voice interfaces, which map sound inputs to an intent, but have a fixed repository of intents and rules for responding to them, resemble the expert systems of old. Instead of automating highly specialized tasks requiring expert knowledge, these new systems automate more everyday things like checking account balances while being packaged in an interface that is perceived to be more intelligent (but is often just as hard-coded).

同样，许多语音接口将声音输入映射到意图，但是具有固定的意图和响应规则的存储库，类似于旧的专家系统。 这些新系统并没有自动执行需要专家知识的高度专门化的任务，而是使更多日常事务自动化，例如在将帐户打包到被认为更智能(但通常是硬编码)的界面中时，检查帐户余额。

In conclusion…

结论…

Relevance: while less often found as standalone products, the ideas behind expert systems have left a lasting impact on software and are still used today.

相关性 ：专家系统背后的思想虽然很少作为独立产品而被发现，但它们对软件产生了持久影响，并且至今仍在使用。

描述性与程序性知识 (Descriptive versus Procedural Knowledge)

Before moving on to neural networks, I’d like to quickly visit a distinction in the types of knowledge we humans can possess: that between descriptive and procedural knowledge.

在进入神经网络之前，我想快速了解一下人类可以拥有的知识类型：描述性知识和程序性知识之间的区别。

Descriptive (also called explicit or declarative) knowledge, takes the form of facts and rules. It can often be represented in words or symbols, and manipulated using logic or some other calculus.

描述性 (也称为显式或声明性 )知识采用事实和规则的形式。 它通常可以用文字或符号表示，并可以使用逻辑或其他微积分进行操作。

Procedural (also called implicit or performative) knowledge, is more about knowing how to do things. It is often associated with concepts such as intuition, muscle memory, and practice.

程序性 (也称为隐式或执行性 )知识，更多是关于知道如何做事的知识。 它通常与直觉，肌肉记忆和练习等概念相关联。

Think about what goes into riding a bike. Sure, there’s knowing that you sit and put your feet on the pedals, and that movement will keep the bike upright. But what’s more important is simply knowing how to ride a bike, which comes with practice. In this way, bicycle riding has more of a procedural component.

想想骑自行车会发生什么。 当然，您会坐下来并把脚放在踏板上，而这种运动会保持自行车直立。 但是，更重要的是简单地知道如何骑自行车，这是实践带来的。 通过这种方式，骑自行车具有更多的程序组成部分。

On the other hand, consider finding the missing angle of a triangle, given some lengths and angles. Here, the process taken to reach the solution is much more explicit, potentially involving trigonometric identities and the like. Every step could be written down on paper if needed. This problem requires more descriptive knowledge.

另一方面，考虑在给定一些长度和角度的情况下找到三角形的缺失角。 在这里，达到解决方案所需的过程更加明确，可能涉及三角恒等式。 如果需要，每个步骤都可以记录在纸上。 这个问题需要更多的描述性知识。

Of course, many complex tasks require a mix of knowledge types, and the distinction is more convenient than it is necessarily indicative about a deep mechanical truth. However, we’ll see that this distinction is useful to keep in mind when thinking about the potential applications for neural networks.

当然，许多复杂的任务需要多种知识类型，并且区分比表示深层机械真相所必需的要方便得多。 但是，我们将在考虑神经网络的潜在应用时记住这一区别很有用。

神经网络和软件2.0 (Neural Networks and Software 2.0)

“It turns out that a large portion of real-world problems have the property that it is significantly easier to collect the data (or more generally, identify a desirable behavior) than to explicitly write the program.” Andrej Karpathy, Software 2.0

“事实证明，现实世界中的大部分问题都具有以下特征：与显式编写程序相比，收集数据(或更笼统地说，确定期望的行为)要容易得多。” Andrej Karpathy， 软件2.0

You may have noticed above that descriptive knowledge is the sort of thing that makes up the knowledge base and inference rules of expert systems.

上面您可能已经注意到， 描述性知识是构成专家系统的知识库和推理规则的一种事物。

How might one build a knowledge base for procedural knowledge? Well, it would be very difficult, especially because even experts in such domains would struggle to precisely express their own expertise.

一个人如何建立程序知识的知识库？ 好吧，这将是非常困难的，尤其是因为即使是此类领域的专家也难以准确表达自己的专业知识。

That is, it would be very difficult using the conventional method of programming.

即，使用常规的编程方法将是非常困难的。

Machine learning models shift the paradigm from one in which the programmer must provide rules and inputs in order to obtain results, to one where they can provide inputs and results to derive rules. That is:

机器学习模型将范式从程序员必须提供规则和输入以获取结果的范式转变为可以提供输入和结果以导出规则的范式。 那是：

Software 1.0: inputs + rules → results

软件1.0：输入+规则→结果

Software 2.0: inputs + results → rules

软件2.0：输入+结果→规则

The ingredients for software 1.0 very much resemble the knowledge base and inference engine required for expert systems. The promise of software 2.0 is that the learned rules can be applied to many new inputs by anyone with the software, without requiring that they themselves have the expertise needed to derive results themselves.

软件1.0的组成部分非常类似于专家系统所需的知识库和推理引擎。 软件2.0的承诺是，使用该软件的任何人都可以将学习到的规则应用于许多新输入，而无需他们本身具有自己获得结果所需的专业知识。

We can see how this suggests a new generation of expert systems, powered by neural networks and deep learning, whose purpose, requirements, and relevance are as follows.

我们可以看到这如何暗示由神经网络和深度学习提供支持的新一代专家系统，其目的，要求和相关性如下。

Purpose: expert systems 2.0 automate decisions and perceptions that normally require specialized training and practice in a domain where expertise is more procedural than descriptive.

目的：专家系统2.0使通常需要专门技能而非描述性知识的领域中需要专门培训和实践的决策和见解自动化。

Requirements: to build such systems requires making a dataset where experts apply their knowledge to make judgements (e.g. authors score pieces of creative writing). To update such systems simply requires updating the dataset.

要求：建立这样的系统需要建立一个数据集，专家可以运用自己的知识来做出判断(例如，作者对创意写作进行评分)。 要更新此类系统，只需要更新数据集即可。

Relevance: expert systems 2.0 stand poised to democratize procedural expertise, with little further technical innovation required.

相关性：专家系统2.0可以使程序专业知识民主化，几乎不需要进一步的技术创新。

Just as with old expert systems, knowledge acquisition still requires the time and attention of domain experts. However, to build a machine learning dataset merely requires them to exercise their craft, rather than deconstruct it. Furthermore, representing this knowledge in a dataset is much easier for an engineer than coming up with methods of encoding and manipulating symbolic knowledge. Finally, updating and improving models can be done more smoothly as the datasets grow and evolve over time.

与旧的专家系统一样，知识获取仍然需要领域专家的时间和精力。 但是，构建机器学习数据集仅要求他们行使自己的技能，而不是对其进行解构。 此外，与提出编码和操纵符号知识的方法相比，对于工程师而言，在数据集中表示该知识要容易得多。 最后，随着数据集随着时间的增长和发展，更新和改进模型可以更顺利地完成。

识别专家系统2.0机会 (Identifying Expert Systems 2.0 Opportunities)

If a task requires expert knowledge of the procedural variety, it is ripe for automation using deep learning. Many creative disciplines fit this description — music, dance, art, etc.

如果一项任务需要专业的程序知识，那么使用深度学习进行自动化的时机已经成熟。 许多创意学科都符合这种描述-音乐，舞蹈，艺术等。

In 2018, Google Arts & Culture Lab made Living Archive, a model trained to predict Wayne McGregor’s dance moves, and thus could be used to suggest choreography.

Google艺术与文化实验室于2018年制作了Living Archive ，该模型经过训练可预测韦恩·麦格雷戈的舞蹈动作，因此可用于建议编舞。

On the technical side, it is far easier to evaluate performances and works than it is to synthesize new ones. Software like SmartMusic is able to score aspects such as the intonation and rhythm of audio recordings. However, these things can be detected and analyzed using hard-coded rules. The real value in areas like music comes when models score more subtle and interpretive elements like dynamics and rubato.

在技​​术方面， 评估性能和工作要比合成新的要容易得多。 像SmartMusic这样的软件能够对录音的音调和节奏等方面进行评分。 但是，可以使用硬编码规则来检测和分析这些事物。 当模型获得诸如动感和鲁巴托之类的更微妙和解释性的元素时，音乐等领域的真正价值就会出现。

If a human expert can reach a decision in a few seconds, before having time to explain how exactly they reached that decision, then capturing many millions of these decisions will bring us towards a machine that can do the same.

如果人类专家可以在几秒钟内做出决定，然后再花时间解释他们如何精确地做出该决定，那么捕获数百万个这样的决定将使我们迈向可以做到这一点的机器。

As is the ideal for AI technology, expert systems 2.0 will augment and democratize human intelligence, rather than make it obsolete. People should still go to their violin or dance lessons every week, but they will also have instant access to valid feedback even when their teachers aren’t in front of them. Shortening the distance to feedback will accelerate learning dramatically.

作为AI技术的理想选择，专家系统2.0将增强人类的智能并使之民主化，而不是使其过时。 人们仍然应该每周都去上小提琴或舞蹈课，但是即使他们的老师不在面前，他们也可以立即获得有效的反馈。 缩短与反馈的距离将大大加快学习速度。

Now, why do I say deep learning based systems will democratize expertise when that hasn’t really been true for the old expert systems? A few factors come into play here:

现在，为什么我说基于深度学习的系统将使专业知识民主化 ，而对于旧的专家系统却并非如此？ 一些因素在这里起作用：

• Old expert systems were often designed in domains where the cost of error was great — law, drug design, medical diagnosis — and therefore there was less value in democratization.
  旧的专家系统通常是在错误成本很高的领域(法律，药物设计，医疗诊断)中设计的，因此民主化的价值较低。

• Since the internet, our culture has shifted towards valuing open access to information.

  自从互联网以来，我们的文化已经转向重视开放信息的获取 。

• Related to the above, access to technology is greater now than ever before.

  与上述相关的是，如今获得技术的机会比以往任何时候都多。

Neural networks and deep learning might just be a stepping stone on the path to artificial general intelligence, but that doesn’t mean that they don’t have immediate value to offer.

神经网络和深度学习可能只是通往人工智能的道路上的垫脚石，但这并不意味着它们没有立即提供的价值。

By automating and democratizing procedural knowledge in the way that expert systems automated descriptive knowledge, neural networks can improve our access to expertise and augment the way we learn forever. Their value has only just begun to be realized.

通过以专家系统自动化描述性知识的方式使过程知识自动化和民主化，神经网络可以改善我们对专业知识的访问，并增加我们永远学习的方式。 它们的价值才刚刚开始被实现。

翻译自: https://towardsdatascience.com/expert-systems-2-0-c8c552f6b2d8

专家库管理软件专家评审系统