Data Scientists use software they didn’t write and data they don’t own pretty much all the time. It is only thanks to open source that they can use programming languages like Python and R, or libraries like Scikit-Learn and TensorFlow, or databases like SQLite and MongoDB. This should not at all be taken for granted. In fact, given how important and ubiquitous intellectual property is in the data science world, it is not being discussed enough I believe. This is why I wrote this blog post.

d ATA科学家们使用的软件，他们没有写，数据不属于他们几乎所有的时间。 只有开源，他们才能使用Python和R这样的编程语言，或者Scikit-Learn和TensorFlow这样的库，或者SQLite和MongoDB这样的数据库。 这根本不应被视为理所当然。 实际上，考虑到知识产权在数据科学世界中的重要性和普遍性，我认为讨论的不够多。 这就是为什么我写这篇博客文章。

I’ll try to answer questions such as:

我将尝试回答以下问题：

• Which intellectual property rights apply to data science?哪些知识产权适用于数据科学？
• Can you copy random public GitHub code, and can others use yours?您可以复制随机的公共GitHub代码，其他人可以使用您的吗？
• Who owns data and the data it is stored in?谁拥有数据以及存储在其中的数据？
• Why did nobody patent neural networks?为什么没有人为神经网络申请专利？

Let’s start with a different question, Why do intellectual property rights even exist? This is not uncontroversial. In fact, there is debate about whether they even should exist to such a large extent.

让我们从另一个问题开始，为什么知识产权甚至存在？ 这并非毫无争议。 实际上，关于它们是否应该在如此大的范围内存在争议。

The main social purpose of the protection of intellectual property is to encourage and reward creative work. It is believed that IP rights stimulate investment (time and money) and lead to innovation.

保护知识产权的主要社会目的是鼓励和奖励创造性的工作。 人们认为，知识产权会刺激投资(时间和金钱)并导致创新。

The three main concepts of intellectual property are copyrights, trademarks, and patents.

知识产权的三个主要概念是版权 ， 商标和专利 。

A copyright grants the author of an original piece of work ownership and exclusive rights for (commercial and non-commercial) distribution. The right exists from the moment of creation, meaning that it need not be registered, although it can be. The work, however, must be fixed in a tangible medium (eg on paper or as bits on a computer), so a thought running through your head is not protected. Besides obviously books, paintings, and music, computer software is also protected by copyright. When it comes to data, things get more complicated (but more that later). Importantly, however, it is only the expression itself that is protected and never the idea behind it. Copyrights are valid until at least 50 years after the death of the author.

版权授予作者原创作品的所有权和(商业和非商业)发行的专有权。 该权利从创建之日起就存在，这意味着尽管可以注册，但无需注册。 但是，必须将作品固定在有形的介质中(例如，固定在纸上或计算机上的碎屑中)，以免影响您的思想。 除了明显的书籍，绘画和音乐外，计算机软件还受版权保护。 当涉及到数据时，事情变得更加复杂(但以后会更加复杂)。 但是重要的是，仅表达式本身受保护，而从不隐藏其背后的思想。 版权有效期至作者去世后至少50年。

A trademark is a sign that distinguishing one company (or goods thereof) from others. An organization’s logo is protected by trademark, but it can also be a phrase like Nike’s Just Do It. There is no registration required for trademarks and they may last indefinitely. It is not of much relevance to data science, however.

商标是将一个公司(或其商品)与其他公司区分开的标志。 组织的徽标受商标保护，但也可以是诸如Nike的Just Do It之类的短语。 商标不需要注册，商标可以无限期使用。 但是，它与数据科学关系不大。

Patents, in contrast, do not exist automatically. They require an application. and unlike copyrights and trademarks, patents protect ideas and inventions rather than the form of expression. Which inventions receive a patent and which not is generally more art than science. A patent grants the owner the exclusive right to commercially exploit the invention, generally for 20 years. A famous patent is Swiffer’s Wet Floor sheet (see the patent application here). Patenting software is possible too, but more on that later.

相反， 专利并不自动存在。 他们需要一个应用程序。 与版权和商标不同，专利保护的是思想和发明，而不是表达形式。 哪些发明获得了专利，哪些通常不是科学。 专利授予所有者专有的商业使用发明的权利，通常为20年。 Swiffer的湿地板是一项著名的专利(请参见此处的专利申请)。 专利软件也是可能的，但以后会更多。

In addition to those three, trade secrets are often also considered a method for IP protection, although not a formal one. Code or data that is not made public may be considered a trade secret.

除了这三个之外， 商业秘密通常也被认为是一种知识产权保护的方法，尽管它不是正式的方法。 未公开的代码或数据可能被视为商业秘密。

开源许可 (Open Source Licensing)

Intellectual property can not only be commercialized by marketing the products they protect but also by giving away the right itself through license agreements. In other words, the owner of IP can confer their right to another party, often in exchange for a royalty fee (or sometimes even for free). Such an agreement can be given exclusively to one other party, or to many (non-exclusive licensing).

知识产权不仅可以通过销售其保护的产品进行商业化，而且可以通过许可协议放弃其本身的权利。 换句话说，IP的所有者可以将自己的权利授予另一方，这通常是为了收取专利使用费(甚至有时是免费的)。 这样的协议可以仅授予另一方，也可以授予许多人(非专有许可)。

In the data science world, licensing is hugely important and basically everywhere. Remember that every piece of software is protected by copyright. That means, principally only their respective authors should be allowed to use each of Python, R, Scikit-Learn, TensorFlow, or any other software. Thankfully, open-source licensing is very common today. For an exact definition of open source, see here. Roughly, it’s three pillars are

在数据科学世界中，许可非常重要，并且基本上在任何地方。 请记住，每个软件都受版权保护。 这意味着，原则上仅应允许其各自的作者使用Python，R，Scikit-Learn，TensorFlow或任何其他软件。 幸运的是，开源许可在今天非常普遍。 有关开源的确切定义，请参见此处 。 大致来说，这是三个Struts

• the source code is publicly accessible源代码可公开访问
• the software can be used for free (there is no royalty)该软件可以免费使用(没有使用费)
• derivative works can be made of it可以制作衍生作品

While everyone can simply write their own open source license for a piece of software, there are a number of standard licenses that are broadly used. The benefit is that there is no (or less) confusion about what rights they include. Open-source licenses generally fall into three categories, depending on the degree of rights the author grants:

尽管每个人都可以简单地为一个软件编写自己的开源许可证，但是有许多标准许可证被广泛使用。 这样做的好处是，它们所包含的权利没有(或更少)混淆。 开源许可证通常根据作者授予的权利程度分为三类：

• Public Domain公共区域
• Permissive Licenses许可许可
• Copyleft LicensesCopyleft许可

公共领域(许可证) (The Public Domain (License))

When a work’s copyright expires, it falls into the public domain, meaning that it has no owner anymore. Anyone and everyone can do with it what they want. There is debate about whether existing copyrights can be transferred into the public domain if an author wants to. Most say no. Hence so-called Public Domain Equivalent Licenses have been designed to grant the same degree of rights as the public domain. These are the most dismissive licenses, as you can imagine. They waive as many rights as possible; for example they usually don’t even require you to attribute the author. The most common public domain equivalent licenses are the Unilicense, the Zero Clause BSD License, the Creative Commons Zero License (CC0), and the Do What the Fuck You Want To Public License (WTFPL). The latter pretty much sums it all up. The only thing you can’t do with such a license is to claim the work is yours since it belongs to the public domain (or kind of). An example of modern software in the public domain is SQLite.

当作品的版权到期时，它属于公共领域，这意味着它不再拥有所有者。 任何人和每个人都可以按照自己的意愿进行操作。 如果作者愿意，是否可以将现有版权转移到公共领域存在争议。 多数人拒绝。 因此，所谓的“ 公共领域等效许可证”已被设计为授予与公共领域相同程度的权利。 您可以想象，这些是最令人讨厌的许可证。 他们放弃了尽可能多的权利。 例如，他们通常甚至不要求您为作者指定属性。 最常见的公共领域等效许可证是Unilicense ， 零条款BSD许可证 ， Creative Commons零许可证 (CC0)和“ 做什么都想让公共许可证 (WTFPL)”。 后者几乎总结了全部。 使用此类许可，您唯一无法做的就是声称该作品属于您，因为它属于公共领域(或某种公共领域)。 SQLite是公共领域中现代软件的一个示例。

Here is the CC0 license text as an example.

这是CC0许可证文本的示例。

CC0 License

CC0许可证

No CopyrightThe person who associated a work with this deed has dedicated the work to the public domain by waiving all of his or her rights to the work worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.

没有版权将作品与该契约相关联的人通过在版权法允许的范围内放弃他或她在全球范围内根据版权法对作品的所有权利，包括所有相关和邻接权，将作品献给了公共领域。

You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission.

您可以复制，修改，分发和执行作品，甚至出于商业目的，而无需征求许可。

许可软件许可 (Permissive Software Licensing)

Licenses that only include minimal restrictions such as an attribution clause are called permissive software licenses. Between the various licenses that fall in this category, there are subtle (and some notable) differences. But roughly said, they all allow you free use and redistribution of the software as long as you attribute the original authors in any derivative work (the attribution clause). However, one major restriction of permissive licenses is that they do not guarantee that future versions of the software remain publicly available, i.e. the software can be later made proprietary by the author. The most common permissive software license is the MIT License. Also very common are the BSD License and the Apache 2.0 License.

仅包括最小限制(例如，归因条款)的许可证称为许可软件许可证。 在属于该类别的各种许可证之间，存在细微(且有些显着)的差异。 但是粗略地说，只要您将原始作者归于任何派生作品中(归因条款)，它们都允许您免费使用和重新分发软件。 但是，许可许可的一个主要限制是它们不保证该软件的将来版本仍可公开获得，即，该软件以后可由作者专有。 最常见的许可软件许可证是MIT许可证 。 BSD许可证和Apache 2.0许可证也很常见。

Much of the software that data scientists use has some sort of permissive open source license, including Python, Julia, PyTorch, and many more. The projects maintained by the Apache Software Foundation deserve particular attention as they include TensorFlow, Hadoop, and Spark. Here is an example.

数据科学家使用的许多软件都有某种允许的开源许可证，包括Python，Julia，PyTorch等。 由Apache Software Foundation维护的项目值得特别关注，因为它们包括TensorFlow，Hadoop和Spark。 这是一个例子。

MIT License

麻省理工学院执照

Copyright © 2020 Andreas Stöffelbauer

版权所有©2020 AndreasStöffelbauer

Permission is hereby granted, free of charge, to any person obtaining a copyof this software and associated documentation files (the “Software”), to dealin the Software without restriction, including without limitation the rightsto use, copy, modify, merge, publish, distribute, sublicense, and/or sellcopies of the Software, and to permit persons to whom the Software isfurnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

特此免费授予获得此软件和相关文档文件(以下简称“软件”)副本的任何人无限制地买卖软件的权利，包括但不限于使用，复制，修改，合并，发布，分发，再许可和/或出售本软件的副本，并允许配备有该软件的人员这样做，但须满足以下条件：上述版权声明和本许可声明应包括在本软件的所有副本或实质部分中软件。

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND […]

该软件按“原样”提供，没有任何形式的保证[…]

Copyleft许可 (Copyleft Licensing)

In contrast to permissive software licenses, so-called copyleft licenses are reciprocal because they require any derivative work to have the same license terms, i.e. to have the very same copyright license. One major motivation could be to ensure that the software remains free, which is not guaranteed under a permissive license. The GNU General Public License (GPT) and the AGPL are the most common copyleft licenses. The most famous software under a copyleft license is Linux.

与许可软件许可证相反，所谓的“ Copyleft”许可证是互惠的，因为它们要求任何派生作品具有相同的许可证条款，即具有完全相同的版权许可证。 一个主要的动机可能是确保软件保持免费，而在许可许可下不能保证这一点。 GNU通用公共许可证 (GPT)和AGPL是最常见的Copyleft许可证。 在copyleft许可下，最著名的软件是Linux。

数据科学家和软件许可 (Data Scientists and Software Licenses)

As you have probably realized by now, you cannot simply copy and use any code you find on GitHub. If there is a license included in a repository, check what you are allowed to do with the code. But if there is no license, the code is still protected under copyright law and you are not allowed to copy it — the fact that the code is public makes no difference. That being said, however, you are not infringing copyright for an insignificant and obvious piece of code; say a simple for-loop; often there is simply just one obvious way to code it. In addition, remember that it is only the code that is protected by copyright, not the idea behind it. So there is nothing preventing you from building a new TensorFlow (as PyTorch did) as long as you do it your own way, for example.

正如您可能已经意识到的那样，您不能简单地复制和使用在GitHub上找到的任何代码。 如果存储库中包含许可证，请检查您是否可以使用该代码。 但是，如果没有许可证，则该代码仍受版权法保护，并且您不得复制该代码-该代码是公开的事实没有任何区别。 话虽这么说，您并没有因为使用无关紧要的明显代码而侵犯版权； 说一个简单的for循环； 通常，只有一种明显的编码方式。 另外，请记住，受版权保护的只是代码，而不是受其保护的想法。 因此，只要您以自己的方式进行操作，就不会阻止您构建新的TensorFlow(如PyTorch所做的那样)。

Usually, data scientists don’t have to be overly careful when it comes to common data science tools. As long as they have some sort of open source license (which they typically do), you are good to go. Of course, there is also proprietary data science software such as MATLAB, SAS, and Tableau, but you won’t find any of those on GitHub anyway. Licensing for proprietary software is typically more individualized.

通常，在涉及通用数据科学工具时，数据科学家不必过于小心。 只要他们拥有某种开源许可证(通常这样做)，您就可以开始使用。 当然，也有专有的数据科学软件，例如MATLAB，SAS和Tableau，但是无论如何您都不会在GitHub上找到任何此类软件。 专有软件的许可通常更加个性化。

How to open source your own code? To open source your own repository or project, you first have to think about which license fits your need. GitHub has a dedicated help page for that (see here). The main step is very easy: all you need to do is include your terms and condition as a license file in your repository. Again, GitHub lets you add the most common licenses with a few clicks. That’s why you’ll find license.md or license.txt files in many GitHub repositories.

如何开源自己的代码？ 要开源自己的存储库或项目的源代码，您首先必须考虑适合您需要的许可证。 GitHub有专门的帮助页面(请参阅此处 )。 主要步骤非常简单：您要做的就是将您的条款和条件作为许可证文件包含在存储库中。 再次，GitHub允许您单击几下即可添加最常见的许可证。 这就是为什么您会在许多GitHub存储库中找到license.md或license.txt文件的原因。

数据科学家的软件专利 (Software Patenting for Data Scientists)

As mentioned previously, copyright only protects the literal expression of a computer program but never the idea behind it, which is often more valuable. Therefore, many firms want the opposite of open source, i.e. to protect their code. A patent would be the logical solution. However, it’s relatively uncommon to patent software. For one thing, think about the fact that a patent requires full disclosure of the source code, which would allow others to reverse-engineer it. In contrast, simply keeping code private (trade secret) is probably more effective than a patent in many cases.

如前所述，版权仅保护计算机程序的字面表达，而不能保护其背后的思想，这通常更有价值。 因此，许多公司都希望与开源相反，即保护其代码。 专利将是合乎逻辑的解决方案。 但是，它相对不常见于专利软件。 一方面，考虑一个事实，即专利要求完全公开源代码，这将使其他人对其进行反向工程。 相反，在许多情况下，仅将代码保密(商业秘密)可能比专利更有效。

What about algorithms such as neural networks? Patenting algorithms is an even more delicate issue and will become more important as AI is making progress. It is not only a question of law but particularly one of ethics.

神经网络之类的算法呢？ 专利算法是一个更加棘手的问题，随着AI的发展，它将变得更加重要。 这不仅是法律问题，而且尤其是道德问题之一。

Principally, facts are not patentable, and since machine learning algorithms like neural networks are basically mathematical methods, they are exempt from protection. However, applied to a certain problem, an algorithm may become part of a patent. So yes, if framed it in the right way, patenting an algorithm is possible. For example, a deep learning algorithm generating a certain kind of audio may be eligible. But that would not prevent the network from being applied to any other problem.

原则上说，事实是不可专利的，并且由于机器学习算法(例如神经网络)基本上是数学方法，因此不受保护。 但是，应用于某个问题的算法可能成为专利的一部分。 因此，是的，如果以正确的方式对其进行构架，就可以申请一种算法。 例如，生成某种音频的深度学习算法可能是合格的。 但这不会阻止网络被应用于其他任何问题。

In addition, I think that the enforcement of software patents is very difficult, and the field of machine learning is moving too fast for many patents to be worth a lengthy application. Perhaps once more powerful AI algorithms arise will patents to play a bigger role. But again, there are ethical questions.

另外，我认为软件专利的执行非常困难，并且机器学习领域的发展速度太快，以至于许多专利不值得长期申请。 也许一旦出现更强大的AI算法，专利将发挥更大的作用。 但是，再次有道德问题。

知识产权与数据 (Intellectual Property and Data)

Data ownership is obviously an important topic for data scientists, but it has not really been discussed a lot. Generally, copyright applies to data. However, data is distinct from software or other creative expressions and therefore deserves a closer look.

数据所有权对于数据科学家来说显然是一个重要的话题，但实际上并没有得到太多讨论。 通常，版权适用于数据。 但是，数据与软件或其他创意表达方式不同，因此值得仔细研究。

To begin with, anything protected by copyright must be creative. That excludes data that simply represent facts, such as data on weather, sports events, or stock prices — perhaps most data that exist are exempt from protection for that reason.

首先，受版权保护的任何内容都必须具有创造力。 这不包括仅表示事实的数据，例如天气，体育赛事或股票价格的数据，因此，可能存在的大多数数据因此不受保护。

In addition to creativity, copyright protection always requires fixation in some tangible medium. With respect to data, this naturally means that datasets and databases play a central role. Indeed, copyrights on a database must be distinguished from copyrights on its content (i.e. the data). It is perfectly imaginable (even common) that the creative way data is combined in a database enjoys copyright protection while the data itself does not. For example, there is a copyright on the famous MNIST dataset (but the author made it available to the public domain).

除创造力外，版权保护始终需要固定在某种有形的媒介上。 对于数据，这自然意味着数据集和数据库起着核心作用。 实际上，必须将数据库的版权与数据库内容(即数据)的版权区分开。 完全可以想象(甚至是很常见)，将数据组合到数据库中的创造性方式享有版权保护，而数据本身却没有。 例如，著名的MNIST数据集上有版权(但作者将其提供给了公共领域)。

However, like with software, the most effective way to protect the intellectual property of data may be simply to keep it a secret. Copyright or not, that means others are prevented from using it. Ethical issues arise again, especially when it comes to personal data.

但是，像使用软件一样，保护数据知识产权的最有效方法可能只是简单地将其保密。 是否拥有版权，这意味着禁止他人使用它。 道德问题再次出现，尤其是在涉及个人数据时。

In short, while most data is not protected by copyright, the way data is organized sometimes is. However, trade secrets probably play a more important role than copyrights when it comes to data.

简而言之，尽管大多数数据不受版权保护，但数据的组织方式有时是受版权保护的。 但是，在涉及数据时，商业秘密可能比版权起更重要的作用。

结论 (Conclusion)

As much as I would love to write more about intellectual property for data scientists, it’s a vast topic and this blog post was only meant as an introduction. Especially the parts on software patenting and data ownership would require much more explanation. Indeed, there’s so much more to say about how IP relates to machine learning and artificial intelligence, for instance. In addition, I didn't talk about whether copyrights and patents apply beyond borders. And I only touched upon some of the ethical issues. These are just some of the topics I would still like to learn more about myself.

尽管我想为数据科学家写更多有关知识产权的文章，但这是一个巨大的话题，这篇博客文章只是作为介绍。 特别是有关软件专利和数据所有权的部分将需要更多解释。 的确，例如，关于IP如何与机器学习和人工智能相关联，还有很多话要说。 此外，我没有谈论版权和专利是否适用于跨国界。 我只谈到了一些道德问题。 这些只是我仍想进一步了解自己的一些主题。