Have you ever wondered how you learn? Most of us learn by working out the pattern and creating a ruleset — which we apply to the next similar situation. What makes you different is the way you learn: how quickly you can analyze the pattern and store/recall the rules when needed. When we learn, we also gain the ability to apply & teach that learning. But what happens when a machine learns?

您是否想过如何学习？ 我们大多数人通过制定模式并创建规则集来学习—我们将其应用于下一个类似情况。 使您与众不同的是学习的方式：可以快速分析模式并在需要时存储/调用规则。 当我们学习时，我们还将获得应用和教授这种学习的能力。 但是当机器学习时会发生什么呢？

In this article, we’ll explore how a machine learns by first exploring basic concepts of learning and then diving deeper into the specifics of the learning.

在本文中，我们将首先探讨学习的基本概念，然后再深入研究学习的具体内容，从而探讨机器如何学习。

None of this will be technical — we’ll avoid jargon, but explain just about enough for you to explore further into the world of Machine Learning & AI.

这些都不是技术性的，我们会避免使用术语，但要向您解释一下足以让您进一步探索机器学习和AI领域。

接下来是什么？ (What comes next?)

If you were shown a sequence of numbers, you can probably guess the next number in the series — depending upon how long the ‘training’ sequence you got and how much time you were given. Simple sequences like 1, 2, 3, 4, 5 easier for you to guess. In some sequences, you can form two rules that determine the next one in the series. However, when things get complicated — sequences such as 1, 1, 2, 3, 5, 8 or 1, 4, 10, 20 — you’ll need more time & numbers to identify the pattern.

如果显示了一个数字序列，您可能可以猜出该序列中的下一个数字-取决于您获得“训练”序列的时间长短和获得的时间。 简单的序列(例如1、2、3、4、5)让您更容易猜测。 在某些序列中，您可以形成两个规则来确定系列中的下一个规则。 但是，当事情变得复杂时(例如1、1、2、3、5、8或1、4、10、20之类的序列)，您将需要更多的时间和数字来识别模式。

Upon further complication, we are left with a seemingly random string of numbers (4, 10, 15, 26, 48, 59) that doesn’t have any underlying pattern — here time cannot help us. We need to know more sequences of that number or change the way we look at it to figure out that pattern.

进一步复杂化之后，我们剩下的看似随机的数字字符串(4、10、15、26、48、59)没有任何潜在的模式-此时此刻我们无能为力。 我们需要知道更多该数字的序列，或者改变我们看待它的方式以找出该模式。

Thus, we need a lot of data to find patterns in seemingly random sequences.

因此，我们需要大量数据才能找到看似随机的序列中的模式。

传统编程机 (Traditional Programmed Machines)

A programmed machine behaves based on the rules that we’ve given them (or programmed into them). When you open Gmail and compose a mail, you’re simply typing in the words ‘How are you’. And this machine is programmed to know that when the user (you) press the keys H, O & W, it is supposed to display the word “HOW”.

编程的机器的行为基于我们赋予它们(或编程到其中)的规则。 打开Gmail并撰写邮件时，您只需输入“您好吗”字样。 并且对本机进行了编程，使其知道用户(您)按H，O和W键时，应该显示单词“ HOW”。

But, if you’ve typed into the mail composer recently, you’ll notice that Gmail gives suggestions:

但是，如果您最近输入了邮件编辑器，则会注意到Gmail 提供了一些建议 ：

How is Gmail able to know that when you type ‘How’, you are going for ‘How are you?’ and not ‘How is it going?’

Gmail如何知道当您键入“如何”时，您将输入“您好吗？”。 而不是“进展如何？”

学习机 (A Learning Machine)

Maybe, Google took all the mail you’ve sent in the past, and saw that when you start your mails with How, you almost always end up typing ‘How are you?’. If they did that, it would look something like this:

也许Google接收了您过去发送的所有邮件，并发现当您使用“如何”开始发送邮件时，几乎总是会键入“您好吗？”。 如果他们这样做，它将看起来像这样：

Here the Output (or the final answer) is fed along with the inputs and the rules are generated. The rule sets are probabilistic — machines cannot be 100% certain of the outcome.

在这里，输出(或最终答案)与输入一起馈入，并生成规则。 规则集是概率性的-机器不能100％确定结果。

You may wonder, what does it do with the rules? Well, it feeds it into another machine that takes in the new inputs and churns out the output:

您可能想知道，这与规则有什么关系？ 好吧，它将它馈送到另一台机器中，该机器接受新的输入并输出输出：

You’ll also notice that there is a loop back to the previous machine on your behaviour — the machine corrects itself when you make changes in your behaviour, adapting to it.

您还会注意到，您的行为有一个循环回前一台机器-当您更改自己的行为以使其适应时，机器会自行纠正。

在我们继续之前... (Before we continue…)

The keen-eyed among you will notice I’ve not touched upon the ‘Unsupervised Learning’ tag below the Learning machine. There are many ways a machine can learn — supervised, unsupervised, reinforcement, etc. And each of those ways helps in the creation of the final rules. You can find more details on the various types of learnings on NVIDIAs blog. It’s written in simple easy to understand way.

你们当中敏锐的眼神会发现，我还没有触及学习机下方的“无监督学习”标签。 机器可以学习很多方法-有监督，无监督，强化等。每种方法都有助于创建最终规则。 您可以在NVIDIA的博客上找到有关各种学习类型的更多详细信息。 它以简单易懂的方式编写。

机器学习的实用眼光 (A Practical Look At Machine Learning)

We’ve explored the way a machine learns, but unless you see it in action, I don’t think we’ll ever appreciate the struggles of the machine. To simplify affairs, I’ve created a machine that has only one goal — reach a particular destination, without dying.

我们已经探索了机器学习的方式，但是除非您看到它的实际应用，否则我认为我们永远不会欣赏机器的挣扎。 为了简化事务，我创建了一台只有一个目标的机器-到达特定的目的地而不会死 。

The overall synopsis is here:

总体简介如下：

Just like how you would attempt to find your way in the dark by trying out a step at a time, this machine sets out to find its way in the dark by spawning 1000s of babies and setting them loose to find their way in the dark. The one that comes the closest did something right, and hence is taken to be the parent of the next generation of the population — each population evolving to get closer to the goal.

就像您一次尝试一步就尝试在黑暗中找到自己的路一样，这台机器开始通过在黑暗中产卵数千个婴儿并将它们放开以在黑暗中找到自己的路来寻找自己的路。 最接近的那个人做了正确的事，因此被认为是下一代人口的父母–每个人口都在进化以接近目标。

Without complicating further, it is sufficient to say that the way the machine learns, differs based on the inputs and the output.

无需赘言，可以说机器的学习方式因输入和输出而异。

For further reading, viewing and exploration, please follow these links:

要进一步阅读，查看和探索，请访问以下链接：

1. Introduction to Machine Learning, MIT OpenCourseware

   MIT OpenCourseware机器学习简介

   (Maths & Technically oriented)

   (数学和技术导向)

2. Learning by Neural Networks by 3Blue1Brown — series of video illustrations on how Neural Networks work. Neural Networks are essentially the techniques machines use to create a ruleset.

   3Blue1Brown的《通过神经网络学习》 —有关神经网络如何工作的一系列视频插图。 神经网络本质上是机器用来创建规则集的技术。

3. (My Favorite & NSFW) How AI Learns by CodeBullet

   (我最喜欢的＆NSFW)CodeBullet如何学习AI

   I love his way of explaining some really complex concepts with a whacky (explicit) easy to understand language. In fact, his evolutionary algorithm example

   我喜欢他用一种怪异(清晰的)易于理解的语言来解释一些非常复杂的概念的方式。 实际上，他的进化算法示例

   I’ve literally copied and modified to simplify some of the concepts.

   我已经进行了字面上的复制和修改，以简化某些概念。

4. Artificial Intelligence and the Future of Business | Hans-Christian Boos | TEDxWHU

   人工智能与商业未来 汉斯·克里斯蒂安·布斯| 泰德

   This will give you a glimpse of where this technology is going(it’s a 2018 video and most of the stuff has already been achieved)

   这将使您了解该技术的发展方向(这是2018年的视频，大部分内容已经实现)

And that’s that! I hope I was able to breakdown the key concept behind learning for machines. Thank you for reading!

就是这样！ 我希望我能够分解机器学习背后的关键概念。 感谢您的阅读！

Checkout more ‘What happens when’ articles here:

在此处查看更多“何时会发生什么”文章：