Principles of Machine Learning -- Before You Start 翻译

全世界都在学习AI，当然我也不能例外。自动驾驶、人脸识别、遍地的机器人。。。So，今天起，我将开始着手翻译Principles of Machine Learning全书，全书共7个章节加一个导读，如果中间掺杂有实验，我也会和大家一起来完成。那么现在，让我们开始机器学习的旅程吧！

Introduction

Welcome to the principles of Machine Learning! My name is Cynthia Rudin. >> And I’m Steve Elston. >> Now machine learning is everywhere. This is the time for machine learning;

it’s becoming mainstream, it’s in the search engines we use every day, it’s in the bank teller machines reading our checks, it’s in our smart phone assistance like Cortana, it’s – you know,

jobs in machine learning are in every industry and we are thrilled to be able to give you an instruction to machine learning in this course. So let’s Steven and I introduce ourselves first.

So I am an associate professor of computer science and electrical and computer engineering at Duke, and an associate professor of statistics at MIT, and my main expertise is in machine learning and data mining.

My lab is called “the prediction analysis” lab. And I have a PhD from Princeton University, and a lot of my work that I do is applied in machine learning and it’s applied to problems in the electric power history, in healthcare,

and in computational criminology. >>

Hi and I’m Steve Elston. I’m a co-founder and principle consultant at a data science consultancy at Seattle called Quantia Analytics. I’ve been working in predictive analytics and machine learning for several decades now.

I’ve been a long-term R S/SPLUS Python user and developer, started using S when it was a Bell labs project and of course more – you know, in recent decade moved to R like everybody else.

I’m currently an advisor on Azure machine learning and some other analytics products to Microsoft, and I’ve worked in a variety of industries:

payment fraud prevention, telecommunication, capital markets including things like market credit risk models, clearing, and collateral management,

and also worked in several industrial areas such as forecasting for logistics management.

And I have a PhD also from Princeton University and mine is in geophysics. >> Now when I first learned about machine learning, I thought it was magic.

A way for computers to predict the future, just by seeing the past. And you know, it’s a way for computers to learn on their own how to solve problems that I can’t solve, and that’s exactly what’s going on.

Computers are learning, just from observing what’s happened in the past. But it’s nothing like magic. Now machine learning, in addition to being a really useful toolbox for industrial applications,

it also gives you a perspective about the way your mind works. So let’s say that I asked you why you could learn and why a computer can’t, right, what would you say?

Would you say that it’s because you’ve seen more of the world than a computer has?

I mean, I think that’s not particularly true anymore, because we have lots of pictures and video and sound now that we could feed to any computer.

Is it because there are more connections in your brain than in a computer? Well that might be part of it, but lots of creatures with much smaller brains than my computer can still learn,

so that’s not it. Maybe you could argue that a brain is more flexible in some ways than a computer; maybe you could think your brain is somehow more open to identifying new types of patterns than your computer,

and that’s why you can learn perhaps.

The interesting thing is that actually that’s not quite the way it is; in fact, it’s sort of the opposite.

Your brain is really good at identifying only certain kinds of patterns; in fact, these are the types of patterns that it’s expecting.

The fact that humans can learn is not so much a consequence of so much of the human brain being flexible, as it is of the human brain being inflexible,

being wired to identify exactly the types of patterns that it comes across, right. Natural images, real sounds, patterns of behavior… these are – you know, these are things that we’re really good at identifying. Humans are absolutely awful at identifying patterns in large databases,right, we can’t – we just can’t learn in some settings, and what enables us to learn in the settings we can learn in is the way that our brains are wired. It’s the structure in our mind; it’s not the flexibility, it’s the limited flexibility.

It’s just that structure. Okay so what is the field of machine learning exactly? It completely revolves around setting up structures in the computer that limit its flexibility and allow it to learn.

Okay, setting up these structures is really a form of statistical modelling, and that’s what we’re going to do in this course. And once you can teach a computer to learn, there are a huge number of applications that you can use it on.

>> So, let’s talk about a few of the applications that we’ll use both for our demos and for the labs that you’re going to do hands-on in this course. So first off, we’re going to do a classification example,

and we’ll be coming back to this in several points in the course – actually each of these and these examples,

and so we’re going to work on classifying diabetes patients who have been in a hospital for treatment and we want to classify the ones who are at high risk that they’re going to be readmitted to the hospital;

that is, that somehow their treatment or the follow up to their treatment or something isn’t likely to be sufficient and they’re going to wind up being re-hospitalized, which is, as you can imagine a serious problem.

It’s expensive, it’s dangerous for the patients, etc. so there’s a lot of reasons why this is an important area. We’re going to look at forecasting; forecasting for demand is used all over the place from warehouse management to power generation.

In particular, we’re going to look at forecasting demand for rented bicycles, and so that will be an – again, an application we’ll come back to at several points in this course.

A lot of these things are done in clustering and segmentation, and we’re going to look at segmenting people by their income level, and that’s an –

again, an analog for lots of different things that are done and everything from political science to marketing. And finally, we’re going to look at how a recommender works;

we’re going to use a restaurant database of Mexican restaurants and compute some recommendations for some of the customers who have written reviews for these Mexican restaurants. >>

Okay now as I mentioned, humans are lousy at finding patterns in large databases, and so here are some of the applications that we’re working on in my lab that use large databases and machine learning,

and in all of these applications, the answer is really in the data. It really is, and by providing the computer with the proper machine learning structure to find important patterns, we can really make headway into societal problems.

For instance, we’ve been looking at power grid failures and personalized advertising, and healthcare applications.

>> So, why would you want to continue with this course? What should you expect to get out of this course? Well first off, it’s going to be a hands-on introduction to machine learning.

We have some great labs laid out here, there’s going to be demos – so you’re going to gain some practical experience at working with data and applying machine learning algorithms of various types to those data.

We’re going to look at actually all the major focus areas in machine learning, so we’ll cover a wide variety of algorithms,

methods and techniques. We’re going to use Azure machine learning quite a bit for demos and for your labs; and why actually we’re doing this, it’s not only a great environment,

but it’s also a great learning environment because a lot of the tedious stuff is kind of taking care for you, so there’s a lot of things you won’t have to spend time when you do your weekly labs.

Nonetheless, we’ll do a significant amount of data cleaning and visualization using R and/or Python, you can pick which path you’re on. So we’ll be working – you can be building some skills with that. And we hope that as you go along here as you work on these examples as you listen to the theory lectures, you start to build some intuition around analytics and machine learning and how it all fits together and mostly given intuition of what’s a useful result, what’s adding value, and what’s going in the direction you or say your boss wants you to go. And we’re going to minimize the math; there’s not going to be any heavy theories, so if you remember a little bit of calculus and some minimal linear algebra, you should be good to go here.

So what are we going to cover specifically in this course? So the first module, we’re going to discuss an introduction to classification, and classification is – in the history of machine learning is kind of where machine learning grew out of largely. 

Then we’re going to talk about regression, and regression is also – many regression methods that are important in machine learning and they have even a much longer history in statistics going back to the late 19th century. 

We’re going to then talk about how do you – once you have improved machine learning models, how do you evaluate the performance? How do you know what to do to improve that performance? 

We’re going to then look at some more modern powerful methods like tree and ensemble learning methods and if you don’t know what that means, stay tuned you’ll find out a lot about it. 

And we’re going to look at optimization-based learning methods such as sport vector machines and neural networks. And we’ll finish up with clustering and recommenders. >> 

So, as you’re taking this course, we hope you will take some steps to get the most out of it to maximize your learning experience.

So overall, think about the fact that this course is going to be over 6 weeks, we have one module per week over those 6weeks so you can kind of plan your time and your work that way. 

For each module, we have lectures, demos, and labs; and the labs derive from the lectures and the demos and they are for you to do on your own to reinforce key learning concepts.

And you’ll perform the labs using – as I already mentioned, Azure machine learning, but also either R or Python, and I suggest you decide if you’re going to use R or Python.

Every lab has the same materials or the same steps in either language; it doesn’t matter in terms of the learning experience. If you’re very ambitious of course you can try both, but for most people just doing one or the other is going to be just great. So some of you want to get the certificate from this course, so what do you need to know? 

First off, you need a 70% score to pass and get the certificate, and that score is divided between assessments at the end of each of the 6 modules, and the final exam. 

So each module assessment that – or all those module assessments together are half your grade, 50% of your grade, and on each question for the assessment, you actually get two tries so if you mess it up the first time don’t panic, 

you get another chance. The other half of your grade is a final exam at the end of the class. This one you only get one try per question, but by then you’ve been through the lectures, you’ve seen all the demos,

and you’ve done all the labs, and so you should – you know, be in a great position to ace that. 

So we hope you get a lot out of this course, and we’re looking forward to presenting it and I think it’s going to be really great informative class to get yourself bootstrapped into the wonderful world of machine learning!

欢迎来到机器学习的原理!我叫辛西娅·鲁丁。>>，我是Steve Elston。>>现在机器学习到处都是。这是机器学习的时代;

机器学习已成为主流，它应用与我们每天使用的搜索引擎中，它在银行柜员机里读取我们的支票，它在我们的智能手机帮助下，像Cortana，它-你知道的，

机器学习应用于我们工作中的每一个行业，我们很高兴能在这门课上给你一个机器学习的指导。让我们先介绍一下我们自己。

我是杜克大学计算机科学与电子计算机工程的副教授，麻省理工学院的统计学副教授，我的主要专长是机器学习和数据挖掘。

我的实验室被称为“预测分析”实验室，我在普林斯顿大学有一个博士学位，我的很多工作都应用了机器学习，比如说电力历史，医疗保健，

在计算犯罪学。> >

大家好，我是史蒂夫·埃尔斯顿。我是西雅图一家名为Quantia Analytics的数据科学咨询公司的联合创始人和首席顾问。我从事预测分析和机器学习已经有几十年了。

我是一个长期的R S/SPLUS Python用户和开发人员，开始使用S时是在贝尔实验室的一个项目中，当然后来-你知道，在最近十年，像其他人一样转移到R。

我现在是Azure机器学习的顾问和微软的其他分析产品，我在很多行业工作过: 例如支付欺诈预防，电信，资本市场包括市场信用风险模型，清算和抵押品管理，并在多个工业领域工作，如物流管理预测。

我有一个普林斯顿大学的博士学位，我主攻地球物理学。>>现在当我第一次学习机器学习的时候，我觉得它很神奇。

一种通过观察过去来预测未来的方法。你知道，这是一种让电脑自己学习如何解决我无法解决的问题的方法，而这正是正在发生的事情。

计算机正在学习，仅仅是通过观察过去发生的事情。但这并不是魔法。现在机器学习，除了作为工业应用的一个非常有用的工具箱，

它也给你一个关于你的思维运作方式的视角。假设我问你为什么你可以去主动学习，电脑却不会主动学习呢，你会说什么?

你会说这是因为你看到的世界比电脑还多吗?

我的意思是，我认为这不再是事实了，因为我们有很多图片，视频和声音，现在我们可以输入进任何电脑。

是因为大脑中的神经网络比电脑的多吗?这可能是其中的一部分，但是很多大脑比我的电脑小的生物仍然可以学习，

所以这不是它。也许你可以认为大脑在某些方面比电脑更灵活;也许你会认为你的大脑比你的电脑更容易识别出新的模式，

这就是为什么你可以学习。

有趣的是，事实并非如此;事实上，这恰恰相反。

你的大脑非常善于识别特定的模式;实际上，这些是它所期望的模式类型。

人类能够学习的事实与其说是由于人类大脑的灵活，不如说是由于人类大脑的灵活性，

通过连线来确定它所遇到的模式的类型。自然图像，真实声音，行为模式这些都是我们非常擅长识别的东西。

人类对于在大型数据库中识别模式是非常可怕的，对吧，我们不能——我们不能在某些环境中学习，在我们可以学习的环境中，让我们学习的是我们的大脑是如何连接的。这是我们头脑中的结构;这不是灵活性，而是有限的灵活性。

这就是结构！那么机器学习的领域到底是什么呢?它完全围绕在计算机中设置结构，限制其灵活性并允许它学习。

好的，建立这些结构实际上是一种统计模型，这是我们在这门课上要做的。一旦你可以教电脑学习，你可以使用大量的应用程序。

>>所以，让我们来讨论一下我们将会用到的一些应用程序，用于我们的演示和实验室，你们将在这门课上亲自动手。首先，我们要做一个分类的例子，

我们将会在课程的几个方面回到这一点——实际上，每一个例子，

所以我们将致力于对那些在医院接受治疗的糖尿病患者进行分类我们想要对那些高危人群进行分类他们将被重新接纳到医院;

也就是说，他们的治疗或后续治疗可能是不够的他们会被重新送进医院，这是一个很严重的问题。

它很贵，对病人来说很危险，所以这是一个重要的领域有很多原因。我们来看看预测;对需求的预测从仓库管理到发电都使用了。

特别地，我们将会看到租赁自行车的需求预测，这将是一个应用，我们将在这门课的几个点上讨论这个应用。

很多事情都是在聚类和分割中完成的，我们会考虑按收入水平细分人们，这是。

再一次，这是一种类似于许多不同事物的模拟，从政治科学到市场营销。最后，我们来看看推荐者是如何工作的;

我们将使用一家墨西哥餐馆的餐馆数据库，并为一些为这些墨西哥餐馆撰写评论的顾客提供一些建议。> >

好了，正如我刚才提到的，人类在大型数据库中发现模式很糟糕，所以这里有一些应用程序我们在我的实验室中使用大型数据库和机器学习，

在所有这些应用中，答案都是在数据中。它确实是，并且通过提供计算机与适当的机器学习结构来寻找重要的模式，我们真的可以在社会问题上取得进展。

例如，我们一直在研究电网故障和个性化广告，以及医疗应用。

>>，你为什么要继续这门课?你希望从这门课中学到什么?首先，这将是机器学习的入门介绍。

我们这里有一些很棒的实验室，会有一些演示，所以你会获得一些实际的经验，在处理数据和将各种类型的机器学习算法应用到这些数据中。

我们将会看到机器学习中所有主要的重点领域，我们将涉及到各种各样的算法，

方法和技术。我们将使用Azure机器来学习一些演示和实验室的知识;为什么我们要这么做，这不仅是一个伟大的环境，但这也是一个很棒的学习环境因为很多繁琐的事情都是为了照顾你，所以你不用花时间在每周的实验室里。尽管如此，我们将使用R和/或Python做大量的数据清理和可视化，您可以选择您所使用的路径。所以我们会工作——你可以用它来建立一些技能。

和我们希望你在这里工作在这些例子中你听理论讲座,你开始建立一些直觉分析和机器学习和如何相互配合,主要是直觉的一个有用的结果,增加价值是什么,什么方向你或你的老板要你去说。我们要最小化数学;

不会有什么大的理论，所以如果你还记得一些微积分和最小的线性代数，应该最好。

那么我们在这门课中具体要讲什么呢?第一个模块，我们将讨论分类的介绍，分类是——在机器学习的历史中，机器学习是在很大程度上产生的。

然后我们将讨论回归，回归也是很多回归方法在机器学习中很重要他们甚至有更长的历史可以追溯到19世纪晚期。

我们接下来要讲的是如何——一旦你有了更好的机器学习模型，你如何评价它的性能?你怎么知道该怎么做才能提高你模型的性能?

我们将会看到一些更现代的强大的方法比如树和集成学习方法。如果你不知道这意味着什么，请继续关注，你会发现很多关于它的东西。

我们将研究基于优化的学习方法，比如运动向量机和神经网络。我们将以聚类和推荐结束。> >

所以，当你学习这门课程的时候，我们希望你能采取一些步骤来最大化你的学习经验。

所以总的来说，考虑到这门课要超过6周的时间，我们每周有一个模块在这6周内所以你可以安排你的时间和你的工作。

对于每个模块，我们都有讲座、演示和实验室;实验室来自于讲课和演示，它们是你自己做的，以强化关键的学习概念。

你将使用-我已经提到过的，Azure机器学习，也可以使用R或Python，我建议你决定是否使用R或Python。

每个实验室在两种语言中都有相同的材料或相同的步骤;就学习经验而言，这无关紧要。如果你雄心勃勃，你当然可以同时尝试这两种方法，但对大多数人来说，只做一件事或另一件事就太好了。你们有些人想从这门课拿到证书，你们需要知道什么?

首先，你需要一个70%的分数才能通过并获得证书，而这个分数在6个模块的末尾和期末考试中分成两部分。

所以每个模块评估——或者所有这些模块的评估都是你的一半分数，50%的分数，在每一个评估的问题上，你实际上得到了两个尝试如果你第一次把它搞砸了不要惊慌，

你得到另一个机会。你们成绩的另一半是期末考试。这个你只需要一个问题，但是到那时你已经通过了讲座，你已经看到了所有的演示，

你已经做了所有的实验，所以你应该——你知道，处于一个非常有利的位置。

所以我们希望你能从这门课中学到很多东西，我们期待着展示它，我认为这将是一个非常好的信息课程，让你自己进入机器学习的奇妙世界!