0:00:00.000,0:00:04.904
What is machine learning? In this video we[br]will try to define what it is and also try

0:00:04.904,0:00:09.520
to give you a sense of when you want to[br]use machine learning. Even among machine

0:00:09.520,0:00:14.252
learning practitioners there isn't a well[br]accepted definition of what is and what

0:00:14.252,0:00:18.926
isn't machine learning. But let me show[br]you a couple of examples of the ways that

0:00:18.926,0:00:23.600
people have tried to define it. Here's the[br]definition of what is machine learning

0:00:23.600,0:00:28.520
does to Arthur Samuel. He defined machine[br]learning as the field of study that gives

0:00:31.037,0:00:33.554
computers the ability to learn without being[br]explicitly programmed. Samuel's claim to

0:00:33.554,0:00:38.452
fame was that back in the 1950's, he wrote[br]a checkers playing program. And the

0:00:38.452,0:00:43.603
amazing thing about this checkers playing[br]program, was that Arthur Samuel himself,

0:00:43.603,0:00:48.268
wasn't a very good checkers player. But[br]what he did was, he had to program for it to play

0:00:48.268,0:00:52.245
10's of 1000's of games against itself.[br]And by watching what sorts of board

0:00:52.245,0:00:56.698
positions tended to lead to wins, and what[br]sort of board positions tended to lead to

0:00:56.698,0:01:00.725
losses. The checkers playing program[br]learns over time what are good board

0:01:00.725,0:01:04.713
positions and what are bad board[br]positions. And eventually learn to play

0:01:04.713,0:01:09.514
checkers better than Arthur Samuel himself[br]was able to. This was a remarkable result.

0:01:09.514,0:01:14.535
Although Samuel himself turned out not to be a[br]very good checkers player. But because the

0:01:14.535,0:01:19.254
computer has the patience to play tens[br]of thousands of games itself. No

0:01:19.254,0:01:24.275
human, has the patience to play that many[br]games. By doing this the computer was able

0:01:24.275,0:01:29.235
to get so much checkers-playing experience that it eventually became a

0:01:29.235,0:01:33.817
better checkers player than Arthur Samuel[br]himself. This is somewhat informal

0:01:33.817,0:01:38.547
definition, and an older one. Here's a[br]slightly more recent definition by Tom

0:01:38.547,0:01:43.607
Mitchell, who's a friend out of Carnegie[br]Mellon. So Tom defines machine learning by

0:01:43.607,0:01:48.819
saying that, a well posed learning problem[br]is defined as follows. He says, a computer

0:01:48.819,0:01:53.843
program is said to learn from experience[br]E, with respect to some task T, and some

0:01:53.843,0:01:58.678
performance measure P, if its[br]performance on T as measured by P improves

0:01:58.678,0:02:03.764
with experience E. I actually think he came[br]up with this definition just to make it

0:02:03.764,0:02:08.346
rhyme. For the checkers playing[br]example the experience e, will be the

0:02:08.346,0:02:13.253
experience of having the program play 10's[br]of 1000's of games against itself. The

0:02:13.253,0:02:17.735
task t, will be the task of playing[br]checkers. And the performance measure p,

0:02:17.735,0:02:22.399
will be the probability that it[br]wins the next game of checkers against

0:02:22.399,0:02:27.157
some new opponent. Throughout these[br]videos, besides me trying to teach you

0:02:27.157,0:02:32.291
stuff, I will occasionally ask you a[br]question to make sure you understand the

0:02:32.291,0:02:36.891
content. Here's one, on top is a[br]definition of machine learning by Tom

0:02:36.891,0:02:42.292
Mitchell. Let's say your email program[br]watches which emails you do or do not flag

0:02:42.292,0:02:47.826
as spam. So in an email client like this[br]you might click this spam button to report

0:02:47.826,0:02:53.263
some email as spam, but not other emails[br]and. Based on which emails you mark as

0:02:53.263,0:02:59.046
spam, so your e-mail program learns better[br]how to filter spam e-mail. What is the

0:02:59.046,0:03:04.290
task T in this setting? In a few seconds,[br]the video will pause. And when it does so,

0:03:04.290,0:03:09.598
you can use your mouse to select one of[br]these four radio buttons to let, to let me

0:03:09.598,0:03:40.190
know which of these four you think is the[br]right answer to this question. That might

0:03:40.190,0:03:45.747
be a performance measure P. And so, our[br]task performance on the task our system's

0:03:45.747,0:03:50.529
performance on the task T, on the[br]performance measure P will improve after

0:03:50.529,0:03:55.957
the experience E. In this class I hope to[br]teach you about various different types of

0:03:55.957,0:04:00.933
learning algorithms. There are several[br]different types of learning algorithms.

0:04:00.933,0:04:05.650
The main two types are what we call[br]supervised learning and unsupervised

0:04:05.650,0:04:10.690
learning. I'll define what these terms[br]mean more in the next couple videos. But

0:04:10.690,0:04:16.028
it turns out that in supervised learning,[br]the idea is that we're going to teach the

0:04:16.028,0:04:20.513
computer how to do something, whereas in[br]unsupervised learning we're going let

0:04:20.513,0:04:25.016
it learn by itself. Don't worry if these[br]two terms don't make sense yet, in the

0:04:25.016,0:04:29.739
next two videos I'm going to say exactly[br]what these two types of learning are. You

0:04:29.739,0:04:34.070
will also hear other buzz terms such as[br]reinforcement learning and recommender

0:04:34.070,0:04:38.621
systems. These are other types of machine[br]learning algorithms that we'll talk about

0:04:38.621,0:04:42.460
later but the two most used types of[br]learning algorithms are probably

0:04:42.460,0:04:46.791
supervised learning and unsupervised[br]learning and I'll define them in the next

0:04:46.791,0:04:51.123
two videos and we'll spend most of this[br]class talking about these two types of

0:04:51.123,0:04:55.720
learning algorithms. It turns out one of[br]the other things we'll spend a lot of time

0:04:55.720,0:05:00.054
on in this class is practical advice for[br]applying learning algorithms. This is

0:05:00.054,0:05:04.444
something that I feel pretty strongly[br]about, and it's actually something that I

0:05:04.444,0:05:08.167
don't know of any other university[br]teaches. Teaching about learning

0:05:08.167,0:05:12.509
algorithms is like giving you a set of[br]tools, and equally important or more

0:05:12.509,0:05:17.616
important to giving you the tools is to[br]teach you how to apply these tools. I like

0:05:17.616,0:05:22.413
to make an analogy to learning to become a[br]carpenter. Imagine that someone is

0:05:22.413,0:05:26.959
teaching you how to be a carpenter and[br]they say here's a hammer, here's a

0:05:26.959,0:05:31.077
screwdriver, here's a saw, good luck.[br]Well, that's no good, right? You, you, you

0:05:31.077,0:05:34.799
have all these tools, but the more[br]important thing, is to learn how to use

0:05:34.799,0:05:38.927
these tools properly. There's a huge[br]difference between, between people that

0:05:38.927,0:05:43.456
know how to use these machines learning[br]algorithms, versus people who don't know

0:05:43.456,0:05:47.626
how to use these tools well. Here in[br]Silicon Valley where I live, when I go

0:05:47.626,0:05:52.328
visit different companies even at the[br]top Silicon Valley companies very often I see

0:05:52.328,0:05:56.428
people are trying to apply machine[br]learning algorithms to some problem and

0:05:56.428,0:06:00.857
sometimes they have been going at it for[br]six months. But sometimes when I look at

0:06:00.857,0:06:05.121
what they're doing I, I, I say, you know,[br]I could have told them like, gee, I could

0:06:05.121,0:06:09.714
have told you six months ago that you[br]should be taking a learning algorithm and

0:06:09.714,0:06:14.470
applying it in like the slightly modified[br]way and your chance of success would have

0:06:14.470,0:06:19.648
been much higher. So what we're going to[br]do in this class is actually spend a lot

0:06:19.648,0:06:23.523
of time talking about how, if you actually[br]tried to develop a machine learning

0:06:23.523,0:06:27.596
system, how to make those best practices[br]type decisions about the way in which you

0:06:27.596,0:06:31.321
build your system so that when you're[br]applying learning algorithm you're less

0:06:31.321,0:06:35.394
likely to end up one of those people who[br]end up pursuing some path for six months

0:06:35.394,0:06:39.373
that, you know, someone else could have[br]figured out it just wasn't gonna work at

0:06:39.373,0:06:43.515
all and it's just a waste of time for six[br]months. So I'm actually going to spend a

0:06:43.515,0:06:47.707
lot of the time teaching you those sorts[br]of best practices in machine learning and

0:06:47.707,0:06:52.052
AI and how to get this stuff to work and[br]how we do it, how the best people do it in

0:06:52.052,0:06:56.143
Silicon Valley and around the world. I[br]hope to make you one of the best people in

0:06:56.143,0:06:59.905
knowing how to design and build serious[br]machine learning and AI systems. So,

0:06:59.905,0:07:04.698
that's machine learning and these are the[br]main topics I hope to teach. In the next

0:07:04.698,0:07:09.023
video, I'm going to define what is[br]supervised learning and after that, what

0:07:09.023,0:07:13.757
is unsupervised learning. And also, start[br]to talk about when you would use each of them.