1
00:00:00,000 --> 00:00:02,202


2
00:00:02,202 --> 00:00:03,160
KEVIN XU: Hey everyone!

3
00:00:03,160 --> 00:00:06,850
Welcome to our
introduction to ML seminar.

4
00:00:06,850 --> 00:00:09,340
I'm Kevin Xu, a
sophomore at the college,

5
00:00:09,340 --> 00:00:11,740
studying computer science and physics.

6
00:00:11,740 --> 00:00:13,990
ZAD CHIN: I'm Zad, I'm a
sophomore at Harvard College,

7
00:00:13,990 --> 00:00:15,640
studying computer science and maths.

8
00:00:15,640 --> 00:00:19,130
KEVIN XU: And we're so excited
that you could join us here today.

9
00:00:19,130 --> 00:00:21,760
We know that we're in the
middle of final test week,

10
00:00:21,760 --> 00:00:23,650
and so everybody's
just a little stressed.

11
00:00:23,650 --> 00:00:27,640
So, hopefully, we can give a
very interesting, and hopefully

12
00:00:27,640 --> 00:00:30,070
fun presentation about
what to look forward

13
00:00:30,070 --> 00:00:32,710
to as you start implementing
your final projects.

14
00:00:32,710 --> 00:00:34,720
So just some logistics first.

15
00:00:34,720 --> 00:00:39,830
We will have some set points where we'll
take a few questions from the audience.

16
00:00:39,830 --> 00:00:43,820
So feel free to type them into
the chat as you think of them,

17
00:00:43,820 --> 00:00:49,330
and somebody will probably read them
to us when we pause for questions.

18
00:00:49,330 --> 00:00:52,450
Yeah, other than that, let's
just get straight into it.

19
00:00:52,450 --> 00:00:55,490


20
00:00:55,490 --> 00:00:59,180
Right so, of course, our seminar
is about machine learning, or ML.

21
00:00:59,180 --> 00:01:02,150
And, so, the first question,
which is not so obvious,

22
00:01:02,150 --> 00:01:03,870
is what is machine learning?

23
00:01:03,870 --> 00:01:08,360
And, so, I'm sure a lot of you have
heard about the developments that

24
00:01:08,360 --> 00:01:12,530
have been in this field, a lot
about neural networking, and perhaps

25
00:01:12,530 --> 00:01:13,580
reinforcement learning.

26
00:01:13,580 --> 00:01:17,780
And a popular topic, of course,
is the game development theory,

27
00:01:17,780 --> 00:01:22,310
where some computers have solved games
through these complicated machine

28
00:01:22,310 --> 00:01:24,350
learning algorithms and neural networks.

29
00:01:24,350 --> 00:01:29,000
And, so, we just want to
give a quick overview of what

30
00:01:29,000 --> 00:01:31,040
exactly falls under machine learning.

31
00:01:31,040 --> 00:01:33,650
And this is actually
a very broad category

32
00:01:33,650 --> 00:01:36,890
that includes a lot of different--

33
00:01:36,890 --> 00:01:38,660
some ideas, and some fields.

34
00:01:38,660 --> 00:01:40,645
And, so, there are three main ones.

35
00:01:40,645 --> 00:01:42,770
There's unsupervised
learning, supervised learning,

36
00:01:42,770 --> 00:01:43,910
and reinforcement learning.

37
00:01:43,910 --> 00:01:47,030
And a lot of the neural
networks that you think about,

38
00:01:47,030 --> 00:01:50,480
when you think about machine
learning, and reinforcement learning,

39
00:01:50,480 --> 00:01:53,450
fall under one of these categories.

40
00:01:53,450 --> 00:01:57,860
But, at the end of the day, ML
is about taking a lot of data

41
00:01:57,860 --> 00:02:01,670
and having the computer, or an
algorithm, or a program, process

42
00:02:01,670 --> 00:02:04,760
the important parts of that data.

43
00:02:04,760 --> 00:02:07,220
Recognize the patterns
in the data, and attempt

44
00:02:07,220 --> 00:02:09,979
to use that data to
generalize to a bigger data

45
00:02:09,979 --> 00:02:11,780
set that you might be given.

46
00:02:11,780 --> 00:02:15,140
So Zad now is going to present a
little bit more about what exactly

47
00:02:15,140 --> 00:02:17,617
these subfields in ML are.

48
00:02:17,617 --> 00:02:20,700
ZAD CHIN: So the first thing we want
to know is about supervised learning,

49
00:02:20,700 --> 00:02:22,283
so what is supervised learning, right?

50
00:02:22,283 --> 00:02:24,560
So, given some labeled
data, say for example,

51
00:02:24,560 --> 00:02:27,930
we have a data set of
cat and dogs pictures.

52
00:02:27,930 --> 00:02:30,380
So, how can a machine learn
to predict the labels that

53
00:02:30,380 --> 00:02:32,130
generalize to unseen data?

54
00:02:32,130 --> 00:02:35,337
So what we do is we put a
data set of cats and dogs

55
00:02:35,337 --> 00:02:37,670
pictures, and then put it
into a machine learning model.

56
00:02:37,670 --> 00:02:40,670
And let the machine learning
model kind of learn itself on how

57
00:02:40,670 --> 00:02:42,742
to differentiate between cats and dogs.

58
00:02:42,742 --> 00:02:45,950
And then we have some testing data set,
we maybe have cats and dogs, or maybe

59
00:02:45,950 --> 00:02:49,220
other images, and the machine would
kind of like label, oh, this is a cat,

60
00:02:49,220 --> 00:02:52,190
or this is a dog, or this
is not a cat or a dog.

61
00:02:52,190 --> 00:02:54,800
So this is kind of like
a supervised machine

62
00:02:54,800 --> 00:02:59,150
learning, whereby we actually label the
data so the machine can learn from it.

63
00:02:59,150 --> 00:03:01,330
So, the next big idea
in machine learning

64
00:03:01,330 --> 00:03:04,600
is also unsupervised learning, where
the machine actually recognizes

65
00:03:04,600 --> 00:03:06,320
the pattern itself from the data.

66
00:03:06,320 --> 00:03:10,150
So like, we don't label anything,
we don't label it's a dog, or a cat,

67
00:03:10,150 --> 00:03:14,110
we just show all the images
to the machine learning,

68
00:03:14,110 --> 00:03:17,520
and the machine will be like, oh this
feature kind of like resemble a cat,

69
00:03:17,520 --> 00:03:18,340
so this is a cat.

70
00:03:18,340 --> 00:03:21,020
And this feature kind of
resembles a dog, so it's a dog.

71
00:03:21,020 --> 00:03:24,150
So that's kind of like a difference
between unsupervised learning

72
00:03:24,150 --> 00:03:25,150
and supervised learning.

73
00:03:25,150 --> 00:03:26,942
And most of the time,
unsupervised learning

74
00:03:26,942 --> 00:03:31,300
includes like, clustering for example,
in a very high dimensional ICU data.

75
00:03:31,300 --> 00:03:33,453
Like who is most likely
to get readmitted,

76
00:03:33,453 --> 00:03:36,370
and what are the features that those
people are going to be readmitted

77
00:03:36,370 --> 00:03:36,870
or not?

78
00:03:36,870 --> 00:03:39,828
So that's kind of like a difference
between supervised and unsupervised

79
00:03:39,828 --> 00:03:40,390
learning.

80
00:03:40,390 --> 00:03:43,420
And next, Kevin is going to talk
more about what is neural network,

81
00:03:43,420 --> 00:03:46,870
and why it important to
unsupervised learning.

82
00:03:46,870 --> 00:03:50,890
KEVIN XU: So, as Zad mentioned, a lot
of the time you're dealing with data

83
00:03:50,890 --> 00:03:54,125
sets that are hugely multidimensional.

84
00:03:54,125 --> 00:03:57,250
And all that means is, there's a lot
of different categories that you could

85
00:03:57,250 --> 00:04:02,260
consider, such as an election, like
the population, the demographics, how

86
00:04:02,260 --> 00:04:04,510
likely the other candidates
are, et cetera, et cetera.

87
00:04:04,510 --> 00:04:06,880
And you just have all of
these categories of data

88
00:04:06,880 --> 00:04:10,780
that you need to somehow compile
together and get some reasonable kind

89
00:04:10,780 --> 00:04:12,970
of prediction as a result.

90
00:04:12,970 --> 00:04:15,550
And this is where neural
networking shines.

91
00:04:15,550 --> 00:04:19,660
And if you continue
researching this field,

92
00:04:19,660 --> 00:04:22,707
and perhaps continue
with this kind of topic,

93
00:04:22,707 --> 00:04:25,040
you're probably going to see
this kind of picture a lot.

94
00:04:25,040 --> 00:04:28,430
And at first this looks
very incomprehensible.

95
00:04:28,430 --> 00:04:31,420
It's just a giant graph with
a bunch of nodes and lines.

96
00:04:31,420 --> 00:04:37,450
But this is actually just a very brief
picture of what neural networking is.

97
00:04:37,450 --> 00:04:40,285
On the left, you can consider these
as input nodes, and on the right

98
00:04:40,285 --> 00:04:42,910
you have output nodes, and in
the middle you have hidden nodes.

99
00:04:42,910 --> 00:04:46,130
But this doesn't really tell you
anything about what this does.

100
00:04:46,130 --> 00:04:49,360
And, so, you can think about
this giant network of graphs,

101
00:04:49,360 --> 00:04:53,620
or however you might imagine it,
as just a black box function.

102
00:04:53,620 --> 00:04:56,317
And this black box function
takes in your input data,

103
00:04:56,317 --> 00:04:58,900
so the different categories,
your multidimensional input data,

104
00:04:58,900 --> 00:05:01,810
and then it spits out
the output data that

105
00:05:01,810 --> 00:05:03,680
should be good enough for a prediction.

106
00:05:03,680 --> 00:05:06,520
So, in the most simplified
case, let's just

107
00:05:06,520 --> 00:05:09,760
talk about a game, such as tic-tac-toe.

108
00:05:09,760 --> 00:05:13,100
And, so, your input data just might
be the state that you're currently in.

109
00:05:13,100 --> 00:05:16,570
So, what squares are filled, whose
turn is it, et cetera, et cetera.

110
00:05:16,570 --> 00:05:19,580
And your output data might
just be one single number,

111
00:05:19,580 --> 00:05:21,700
the heuristic between zero and one.

112
00:05:21,700 --> 00:05:26,200
And all this number tells you is how
good your current board position is.

113
00:05:26,200 --> 00:05:29,870
So obviously, if you can make
three in a row in your next turn,

114
00:05:29,870 --> 00:05:32,215
your heuristic should be
very good, it should be one.

115
00:05:32,215 --> 00:05:34,840
And if you're going to lose in
the next turn it should be zero.

116
00:05:34,840 --> 00:05:39,730
So, obviously, not every
state falls under this case.

117
00:05:39,730 --> 00:05:42,600
So, there's a continuous range
of numbers between zero and one

118
00:05:42,600 --> 00:05:48,360
that your black box function should
be able to retrieve, or accurately

119
00:05:48,360 --> 00:05:50,340
predict from this input data.

120
00:05:50,340 --> 00:05:53,070
And this is a job of everything
in the middle, right?

121
00:05:53,070 --> 00:05:55,410
How do these variables
relate, what function

122
00:05:55,410 --> 00:05:59,870
or what functions you should apply
to these variables in the middle

123
00:05:59,870 --> 00:06:03,460
of processing, et cetera, et cetera.

124
00:06:03,460 --> 00:06:06,670
And, so, the overall
goal of neural networking

125
00:06:06,670 --> 00:06:09,640
is to design this
middle function, right?

126
00:06:09,640 --> 00:06:15,160
How do you get the computer to
find these patterns, and to make--

127
00:06:15,160 --> 00:06:18,200
create this black box function
in a reasonable manner.

128
00:06:18,200 --> 00:06:20,770
And this is something that's
really hard for humans to do,

129
00:06:20,770 --> 00:06:22,240
because there's so much data.

130
00:06:22,240 --> 00:06:25,592
And, so, we give it to
the machine instead.

131
00:06:25,592 --> 00:06:27,800
ZAD CHIN: So, a general
outline in a machine learning

132
00:06:27,800 --> 00:06:30,060
project, or machine learning
research, for example,

133
00:06:30,060 --> 00:06:31,650
we start with a data collection.

134
00:06:31,650 --> 00:06:34,970
So there are a lot of different
ways where we can collect data.

135
00:06:34,970 --> 00:06:37,220
Some significant ways like
finding sources of data

136
00:06:37,220 --> 00:06:41,790
is like, scrapping it from websites,
API call, or readily available data

137
00:06:41,790 --> 00:06:47,000
sets such as Kaggle, or
like in UCI ML repository.

138
00:06:47,000 --> 00:06:50,720
But do take a note that scrapping
raw data, for example, from API call,

139
00:06:50,720 --> 00:06:53,030
or even websites, may take
a long time to clean up

140
00:06:53,030 --> 00:06:55,860
the data, especially when there
is a lot of empty rows and stuff.

141
00:06:55,860 --> 00:06:58,220
So the next step we do is
normally data exploration.

142
00:06:58,220 --> 00:07:01,340
And initial data exploration
is normally very powerful,

143
00:07:01,340 --> 00:07:04,640
and it gives you great insight
on what features to take in,

144
00:07:04,640 --> 00:07:07,490
what features to drop,
whether your data is biased.

145
00:07:07,490 --> 00:07:11,900
And understanding the dimension of
the data is actually very important.

146
00:07:11,900 --> 00:07:14,570
So the third one is, of
course, choosing an ML model.

147
00:07:14,570 --> 00:07:17,030
Choose a simple model to
start with, say for example,

148
00:07:17,030 --> 00:07:19,670
what is the type of problem
you want to solve, right?

149
00:07:19,670 --> 00:07:22,400
We'll go too deep in
that-- in an example

150
00:07:22,400 --> 00:07:25,010
later, but like, some
problems I consider,

151
00:07:25,010 --> 00:07:27,612
is it a regression or
classification problem?

152
00:07:27,612 --> 00:07:30,320
Or whether to actually use the
supervised or unsupervised machine

153
00:07:30,320 --> 00:07:31,190
learning model.

154
00:07:31,190 --> 00:07:33,240
There are some models
that you can choose from.

155
00:07:33,240 --> 00:07:35,115
And the fourth one,
after you choose a model,

156
00:07:35,115 --> 00:07:38,620
and code it out with all the
libraries, you get the result.

157
00:07:38,620 --> 00:07:39,620
You want to test it out.

158
00:07:39,620 --> 00:07:42,110
What is the accuracy
score of your model,

159
00:07:42,110 --> 00:07:45,350
what is the precision, what is
the record score of your model.

160
00:07:45,350 --> 00:07:49,160
And how can we improve the model
by fine tuning the parameters

161
00:07:49,160 --> 00:07:51,338
or even using the more
sophisticated models?

162
00:07:51,338 --> 00:07:53,630
Say, for example, you started
with logistic regression,

163
00:07:53,630 --> 00:07:55,770
maybe you can move on
to neural networks.

164
00:07:55,770 --> 00:08:01,730
So, I think we can start with a bit
of questions if you guys have any.

165
00:08:01,730 --> 00:08:05,780
Or not, we can move on
to examples that we have.

166
00:08:05,780 --> 00:08:09,290
SPEAKER: OK, there's one
question, from Prateek.

167
00:08:09,290 --> 00:08:13,580
"Is there any difference between
data science and machine learning?"

168
00:08:13,580 --> 00:08:17,040


169
00:08:17,040 --> 00:08:18,152
ZAD CHIN: So--

170
00:08:18,152 --> 00:08:19,520
KEVIN XU: Yeah, do you
want to take this, Zad?

171
00:08:19,520 --> 00:08:20,395
ZAD CHIN: Yeah, sure.

172
00:08:20,395 --> 00:08:22,160
I think data science is more--

173
00:08:22,160 --> 00:08:24,553
I think machine learning
is under data science.

174
00:08:24,553 --> 00:08:25,970
You can correct me when I'm wrong.

175
00:08:25,970 --> 00:08:30,620
I think, like, data scientists somehow
also use machine learning model

176
00:08:30,620 --> 00:08:32,549
to help them to analyze the data.

177
00:08:32,549 --> 00:08:35,450
So, basically, what it means
for data science is that like,

178
00:08:35,450 --> 00:08:41,510
we try to understand the pattern,
or what useful information

179
00:08:41,510 --> 00:08:43,789
can a huge amount of
data tell us, right?

180
00:08:43,789 --> 00:08:47,190
So, I feel like, machine learning
model can be a very good stepping stone

181
00:08:47,190 --> 00:08:48,380
into data science.

182
00:08:48,380 --> 00:08:51,380
And also, there are statistics
as well, like pure statistics,

183
00:08:51,380 --> 00:08:56,225
such as chi-square test, which is
not really machine learning but also

184
00:08:56,225 --> 00:08:58,085
a part of data science.

185
00:08:58,085 --> 00:09:03,320
KEVIN XU: Right, just echoing
what Zad said, ML is a technique.

186
00:09:03,320 --> 00:09:08,090
Or a-- you can always think
of it as a methodology

187
00:09:08,090 --> 00:09:12,960
to approach the study of how
you can extrapolate data.

188
00:09:12,960 --> 00:09:17,827
And, so, there's a lot that falls under
this, that falls on machine learning.

189
00:09:17,827 --> 00:09:19,910
But there are also a lot
of things that you can do

190
00:09:19,910 --> 00:09:22,800
that are not machine learning, right?

191
00:09:22,800 --> 00:09:25,700
And, so, I hope that
answers your question.

192
00:09:25,700 --> 00:09:28,025
SPEAKER: OK, no other
questions, you can continue.

193
00:09:28,025 --> 00:09:30,890


194
00:09:30,890 --> 00:09:34,047
KEVIN XU: Zad will have to [? click. ?]

195
00:09:34,047 --> 00:09:36,500
ZAD CHIN: So you guys
can go to this link.

196
00:09:36,500 --> 00:09:40,130
If you want to make a copy of the
notebook yourself and follow up when we

197
00:09:40,130 --> 00:09:46,290
actually present it, you can go
to this tinyurl.com/ML-notebook-1.

198
00:09:46,290 --> 00:09:49,380
We will jump straight to the
notebook that we can show you.

199
00:09:49,380 --> 00:09:53,240
So there are a few contents, if
you want this table of contents.

200
00:09:53,240 --> 00:09:57,090
And I'll close it, so we can
have a better view of [? it. ?]

201
00:09:57,090 --> 00:09:58,470
So let's get started.

202
00:09:58,470 --> 00:10:02,940
We talked a lot about how we can
start with machine learning research.

203
00:10:02,940 --> 00:10:05,780
So, in this particular
notebook, we will try

204
00:10:05,780 --> 00:10:08,660
to use the Iris data set, one
of the most famous data sets

205
00:10:08,660 --> 00:10:10,200
about machine learning.

206
00:10:10,200 --> 00:10:12,710
So, in a supervised
learning, we will try

207
00:10:12,710 --> 00:10:16,940
to use two different models,
which is k-nearest neighbor

208
00:10:16,940 --> 00:10:18,140
and logistic regression.

209
00:10:18,140 --> 00:10:20,420
And, the second part, we
will jump into reinforcement

210
00:10:20,420 --> 00:10:22,680
learning with a different example.

211
00:10:22,680 --> 00:10:26,660
So, as we talked about last
time, the ML development process

212
00:10:26,660 --> 00:10:31,160
includes data collection, train-test
split, identify correct ML problems

213
00:10:31,160 --> 00:10:34,880
and ML model to use, evaluate
the performance of the algorithm,

214
00:10:34,880 --> 00:10:38,270
and try different models
and retrain to retest.

215
00:10:38,270 --> 00:10:39,682
So let's get started.

216
00:10:39,682 --> 00:10:42,890
The first thing you want to do is, of
course, import the necessary libraries,

217
00:10:42,890 --> 00:10:46,920
like Pandas, NumPy, plotting
library, train-test library.

218
00:10:46,920 --> 00:10:49,400
So you get straight and run this.

219
00:10:49,400 --> 00:10:53,840
So to run a cell in Jupyter
Notebook or like, Google Colab,

220
00:10:53,840 --> 00:10:57,257
you just press Shift-Enter.

221
00:10:57,257 --> 00:10:59,090
And, so, the next thing
I'll do is to import

222
00:10:59,090 --> 00:11:01,700
the data set that we talked
about, which is the Iris data set.

223
00:11:01,700 --> 00:11:05,540
So this DataFrame.head is to
print out like-- or how the--

224
00:11:05,540 --> 00:11:08,640
it's to give you an idea on how
the DataFrame actually looks like.

225
00:11:08,640 --> 00:11:14,435
So we can see that if we run this, there
is sepal length, and then petal length

226
00:11:14,435 --> 00:11:15,935
and petal width, and there's target.

227
00:11:15,935 --> 00:11:19,640
Target is the three different things
that we have, which is, I think,

228
00:11:19,640 --> 00:11:22,140
different kinds of iris.

229
00:11:22,140 --> 00:11:24,660
And the next thing we want
to do is try to explore

230
00:11:24,660 --> 00:11:27,510
what is the shape of the
DataFrame, for example, right?

231
00:11:27,510 --> 00:11:31,380
Now, we know that there are 150
rows and five different columns,

232
00:11:31,380 --> 00:11:35,252
here we can see there's five different
columns, and there's 150 rows.

233
00:11:35,252 --> 00:11:37,210
So, first of all, the
first thing we want to do

234
00:11:37,210 --> 00:11:40,430
is to actually explore
and analyze the data.

235
00:11:40,430 --> 00:11:44,320
That's because it's important to do
initial data exploration because it

236
00:11:44,320 --> 00:11:46,570
might be biased, it
might be noisy, and it

237
00:11:46,570 --> 00:11:49,210
shows the relationship between
different features and target

238
00:11:49,210 --> 00:11:50,830
and better train our ML.

239
00:11:50,830 --> 00:11:54,160
So I put some links here
that you can use next time.

240
00:11:54,160 --> 00:11:56,860
But we will start by
using-- putting a pie

241
00:11:56,860 --> 00:11:58,690
chart between the different targets.

242
00:11:58,690 --> 00:12:03,580
Like, there's setosa,
virginica, and versicolor.

243
00:12:03,580 --> 00:12:05,620
That's a different type of iris.

244
00:12:05,620 --> 00:12:07,480
I can see that each
of them, kind of like,

245
00:12:07,480 --> 00:12:12,130
33%, represented very equally and
balanced in the labeled data set.

246
00:12:12,130 --> 00:12:18,160
And this is to plot the number
of different labels versus--

247
00:12:18,160 --> 00:12:22,120
this is the number of like--
count, and then this is the width.

248
00:12:22,120 --> 00:12:24,130
So you can see that most of the--

249
00:12:24,130 --> 00:12:29,350
say for example, sepal, we can
see fall in the 3.0 to 3.5,

250
00:12:29,350 --> 00:12:33,860
and it's kind of evenly,
or normal distributed.

251
00:12:33,860 --> 00:12:38,590
And this is more like a graph to see
how different kinds of length and width

252
00:12:38,590 --> 00:12:41,590
vary according to species.

253
00:12:41,590 --> 00:12:44,380
And I take some time, and I
think this is very useful.

254
00:12:44,380 --> 00:12:46,750
And this is basically
a pairwise plotting

255
00:12:46,750 --> 00:12:50,710
between how the features and
the target are influenced.

256
00:12:50,710 --> 00:12:54,370
It takes some time to
plot, but it will be cool.

257
00:12:54,370 --> 00:12:57,580
So say, for example, look at this.

258
00:12:57,580 --> 00:13:01,220
We can see that like, petal length has
a very huge influence on the target.

259
00:13:01,220 --> 00:13:04,220
You can see that for target
zero, they have really,

260
00:13:04,220 --> 00:13:07,060
really, significant smaller
petal length, as opposed

261
00:13:07,060 --> 00:13:10,630
to target two which actually has
a very clear distinct feature.

262
00:13:10,630 --> 00:13:13,990
And moving forward, to actually--
you know like sometimes,

263
00:13:13,990 --> 00:13:15,830
when you look at graph,
it's not intuitive?

264
00:13:15,830 --> 00:13:20,620
So a very, very good way
to plot the intuition--

265
00:13:20,620 --> 00:13:23,980
to calculate, to see the relationship
between different features,

266
00:13:23,980 --> 00:13:26,692
is to plot a heat map.

267
00:13:26,692 --> 00:13:28,400
You can see that when
we plot a heat map,

268
00:13:28,400 --> 00:13:31,760
you can see the higher correlation
between petal length and target.

269
00:13:31,760 --> 00:13:34,480
This is where we want the label,
right, petal length and target,

270
00:13:34,480 --> 00:13:36,850
and petal width and target.

271
00:13:36,850 --> 00:13:41,680
So we pass it into the machine learning
model where you want to-- we can also

272
00:13:41,680 --> 00:13:43,750
drop-- we can actually
drop, which means delete,

273
00:13:43,750 --> 00:13:47,380
the column of sepal width and length,
because it doesn't really affect--

274
00:13:47,380 --> 00:13:51,208
it doesn't really correlate to the
target that we actually care about.

275
00:13:51,208 --> 00:13:53,500
So the next thing we want to
do is to train-test split.

276
00:13:53,500 --> 00:13:55,960
The importance of train-test split
can be learned more about here,

277
00:13:55,960 --> 00:13:57,110
but we will just move on.

278
00:13:57,110 --> 00:13:59,860
So, basically, when
we have a data set, we

279
00:13:59,860 --> 00:14:02,290
can't just put everything
in your training sample.

280
00:14:02,290 --> 00:14:07,870
We need to split it into maybe 70/30
split, which is 70% training, then

281
00:14:07,870 --> 00:14:12,380
30% testing, or 80/20, 80% testing and--

282
00:14:12,380 --> 00:14:15,405
I'm sorry, 80% training and 20% testing.

283
00:14:15,405 --> 00:14:17,530
Depending on the size of
your data set, some people

284
00:14:17,530 --> 00:14:22,090
go with 70/30, if you have a huge data
set you can actually do 80/20 split.

285
00:14:22,090 --> 00:14:25,090
So, in here, because we
don't have a huge data set,

286
00:14:25,090 --> 00:14:28,270
you only have like
1,500 rows, so I decided

287
00:14:28,270 --> 00:14:35,260
to go with 70/30 split, which we
can see that the test size is 0.3.

288
00:14:35,260 --> 00:14:40,900
And we have, after splitting, we get
105 rows in the training data set,

289
00:14:40,900 --> 00:14:44,830
and 45 rows in the testing data set.

290
00:14:44,830 --> 00:14:48,700
And we can also check whether we
actually trained it correctly,

291
00:14:48,700 --> 00:14:52,150
and this is a very important
step whereby we actually split.

292
00:14:52,150 --> 00:14:54,820
So, now, we have five
different columns, right?

293
00:14:54,820 --> 00:14:57,850
We have sepal length, sepal width,
petal length, and petal width.

294
00:14:57,850 --> 00:15:01,040
And we have target, and this
is what we want to test it out.

295
00:15:01,040 --> 00:15:05,500
So we need to split the training data
set into both x and y, where we have--

296
00:15:05,500 --> 00:15:09,312
these are call features, and then
this is the target that we want.

297
00:15:09,312 --> 00:15:11,770
And these are all features that
are in the testing data set

298
00:15:11,770 --> 00:15:13,760
and this is the target.

299
00:15:13,760 --> 00:15:19,150
So we split that, all
of this can be fine--

300
00:15:19,150 --> 00:15:22,900
this is actually under Pandas, if this
notation seems really, really weird

301
00:15:22,900 --> 00:15:24,280
to you.

302
00:15:24,280 --> 00:15:27,650
It is very well documented
in the Panda DataFrame.

303
00:15:27,650 --> 00:15:31,750
And this is still head, and we
see how it looks like in this

304
00:15:31,750 --> 00:15:36,610
or-- after I split it, you can see
that train x now has four things.

305
00:15:36,610 --> 00:15:40,720
Four different features that we want
to train it on the head, which is this

306
00:15:40,720 --> 00:15:45,040
is the attribute of it, which
correspond to the ID features.

307
00:15:45,040 --> 00:15:49,730
So, next, what you want to do is
identify correct problem and ML model.

308
00:15:49,730 --> 00:15:52,490
So in this problem it's more
like a classification problem.

309
00:15:52,490 --> 00:15:55,630
So, for example, if you are provided
with several inputs and variables,

310
00:15:55,630 --> 00:15:58,330
a classification model will
actually try to predict

311
00:15:58,330 --> 00:16:02,595
the class of the kind of data.

312
00:16:02,595 --> 00:16:04,720
Say, for example, here is
a classification problem,

313
00:16:04,720 --> 00:16:08,870
because we are given the features
sepal length, sepal width, petal length

314
00:16:08,870 --> 00:16:11,410
and petal width, and we want
to find what target, which

315
00:16:11,410 --> 00:16:13,900
is what iris species it is.

316
00:16:13,900 --> 00:16:16,420
So that's a classification problem.

317
00:16:16,420 --> 00:16:19,420
As opposed to a regression problem,
where we are given certain features.

318
00:16:19,420 --> 00:16:22,990
You want to find a petal length or
petal width, that's a regression model.

319
00:16:22,990 --> 00:16:27,470
You can read more about it here, I'm
not going through it for time's sake.

320
00:16:27,470 --> 00:16:30,220
And before we move on
to test out that model,

321
00:16:30,220 --> 00:16:34,850
I want to talk about evaluation metrics
so we can keep running and see it.

322
00:16:34,850 --> 00:16:39,190
So there are a few eval-- before
we dive into evaluation metrics,

323
00:16:39,190 --> 00:16:41,950
we need to understand
what is a true false--

324
00:16:41,950 --> 00:16:45,890
true positive, true negative, and
false positive, and false negative.

325
00:16:45,890 --> 00:16:50,110
So a true positive is when outcome
correctly predicts the positive class.

326
00:16:50,110 --> 00:16:54,280
Say, for example, you have a one,
and then the model actually correctly

327
00:16:54,280 --> 00:16:56,320
predicts the one class.

328
00:16:56,320 --> 00:17:00,250
And, similarly, a true negative is
an outcome where the model actually

329
00:17:00,250 --> 00:17:02,830
correctly predicts the negative class.

330
00:17:02,830 --> 00:17:03,980
Let me give you an example.

331
00:17:03,980 --> 00:17:08,890
So, basically, say we have a cat
data set and a not cat data set.

332
00:17:08,890 --> 00:17:14,829
So a true positive is whereby a model
correctly predicts that it is a cat.

333
00:17:14,829 --> 00:17:17,950
A true negative is when the
model actually correctly predicts

334
00:17:17,950 --> 00:17:19,609
that it is not a cat.

335
00:17:19,609 --> 00:17:22,329
So that's the difference between
true positive and true negative.

336
00:17:22,329 --> 00:17:25,510
A false positive is an outcome
where the model incorrectly

337
00:17:25,510 --> 00:17:26,930
predicts the positive class.

338
00:17:26,930 --> 00:17:29,620
Say, for example, I
have picture of a cat,

339
00:17:29,620 --> 00:17:32,680
but then my model predicts
it as not a cat [? view. ?]

340
00:17:32,680 --> 00:17:34,190
That's a false positive.

341
00:17:34,190 --> 00:17:36,940
And a false negative is an outcome
where the model incorrectly

342
00:17:36,940 --> 00:17:38,750
predicts a negative class.

343
00:17:38,750 --> 00:17:39,820
That's when I give--

344
00:17:39,820 --> 00:17:44,270
I feed in a non-cat picture, but
the model says, this is a cat.

345
00:17:44,270 --> 00:17:47,660
So that's a false negative.

346
00:17:47,660 --> 00:17:49,420
So in terms of a
different kind of metrics

347
00:17:49,420 --> 00:17:53,170
that we use in judging the
performance of an ML metrics,

348
00:17:53,170 --> 00:17:55,940
we have accuracy, recall, precision.

349
00:17:55,940 --> 00:17:58,990
And we also have other called
the ROC which is receiver

350
00:17:58,990 --> 00:18:00,790
operating characteristic curve.

351
00:18:00,790 --> 00:18:04,300
And AUC, which most
often used as opposed

352
00:18:04,300 --> 00:18:07,310
to ROC, which is area under the curve.

353
00:18:07,310 --> 00:18:11,200
So accuracy bias' meaning is the
fraction of predictions that got right.

354
00:18:11,200 --> 00:18:17,870
Say, let's take the example of the
cat and non-cat data set, right?

355
00:18:17,870 --> 00:18:19,750
So the fraction of--

356
00:18:19,750 --> 00:18:24,070
out of the 150 prediction,
for example, if my model

357
00:18:24,070 --> 00:18:28,260
predicts 96% of the cat pictures
correctly, that's an accuracy.

358
00:18:28,260 --> 00:18:32,400
And, also, you have recall, which is
what proportion of actual positive

359
00:18:32,400 --> 00:18:33,360
are defined correctly.

360
00:18:33,360 --> 00:18:37,950
So mathematically, it's defined as
true positive or true positive plus

361
00:18:37,950 --> 00:18:38,850
false negative.

362
00:18:38,850 --> 00:18:43,890
And precision is a true positive or true
positive plus false positive, sorry.

363
00:18:43,890 --> 00:18:46,620
So, it's like, what proportion
of positive identification

364
00:18:46,620 --> 00:18:47,820
were actually correct.

365
00:18:47,820 --> 00:18:50,940
And I thought it might
be a bit overwhelming

366
00:18:50,940 --> 00:18:53,940
to hear a lot like, true positive,
true negative, it's like,

367
00:18:53,940 --> 00:18:56,670
oh my god, this is so much.

368
00:18:56,670 --> 00:19:00,960
I included a lot of different resources
that you can actually go and read it.

369
00:19:00,960 --> 00:19:04,290
I think they are pretty good resources
whereby you can learn more, and get

370
00:19:04,290 --> 00:19:06,730
accustomed to all of these terms.

371
00:19:06,730 --> 00:19:10,713
And, so, let's first dive in
to the first machine learning--

372
00:19:10,713 --> 00:19:12,880
the supervised model that
we're going to talk about,

373
00:19:12,880 --> 00:19:14,800
which is k-nearest neighbor.

374
00:19:14,800 --> 00:19:18,220
So the k, in k-nearest neighbor,
is the nearest neighbor

375
00:19:18,220 --> 00:19:19,830
we wish to take the vote from.

376
00:19:19,830 --> 00:19:21,060
Let's do an example, right?

377
00:19:21,060 --> 00:19:25,995
So this is a dot I want to label,
whether it's a blue or a red dot.

378
00:19:25,995 --> 00:19:30,840
In the radius of three, which
is this-- what's it called--

379
00:19:30,840 --> 00:19:33,300
solid line circle, you
can see that we have

380
00:19:33,300 --> 00:19:38,350
two red triangles but one blue square.

381
00:19:38,350 --> 00:19:43,050
So, in this case, this dot will
be classified by the KNN label

382
00:19:43,050 --> 00:19:45,600
as the red triangle.

383
00:19:45,600 --> 00:19:47,680
And what about if it's larger, right?

384
00:19:47,680 --> 00:19:50,060
So what about k value of five?

385
00:19:50,060 --> 00:19:52,270
So it would be the dotted line circle.

386
00:19:52,270 --> 00:19:54,300
So you can see that in
this dotted line circle,

387
00:19:54,300 --> 00:20:02,700
we can see that if we classify this
green dot, we have three blue squares,

388
00:20:02,700 --> 00:20:04,360
but we have two red triangles.

389
00:20:04,360 --> 00:20:07,980
So this blue dot, based on
this dotted line circle,

390
00:20:07,980 --> 00:20:12,450
will be classified as a blue square.

391
00:20:12,450 --> 00:20:18,110
So we can see that actually choosing
a k value in the KNN algorithm

392
00:20:18,110 --> 00:20:18,860
is very important.

393
00:20:18,860 --> 00:20:21,110
And that's where all the
time consuming parts come in,

394
00:20:21,110 --> 00:20:24,560
because we want to choose the correct
model that doesn't overfit or underfit

395
00:20:24,560 --> 00:20:26,090
the algorithm.

396
00:20:26,090 --> 00:20:30,650
So some of the ways I can do it is plot
a graph of accuracy versus k value,

397
00:20:30,650 --> 00:20:33,020
or graph of error rate versus k value.

398
00:20:33,020 --> 00:20:37,430
But, most of the time, we just test with
a random one and we move on from there.

399
00:20:37,430 --> 00:20:40,708
And there's also a different
metric system that we can use.

400
00:20:40,708 --> 00:20:43,250
If we learn maths we also learn
about the Euclidean distance,

401
00:20:43,250 --> 00:20:46,760
whereby it's like, draw a triangle,
and we have the Euclidean distance.

402
00:20:46,760 --> 00:20:48,530
Or we have the Manhattan
distance, whereby

403
00:20:48,530 --> 00:20:52,430
it's like the longest distance
from a point to a point.

404
00:20:52,430 --> 00:20:57,530
So the KNN resources that I found
really, really helpful in explaining

405
00:20:57,530 --> 00:20:58,980
will be here as well.

406
00:20:58,980 --> 00:21:01,790
So, without further ado, let's
get started with model testing.

407
00:21:01,790 --> 00:21:05,210
So I use a SKlearn model to--

408
00:21:05,210 --> 00:21:09,830
I paste in my k-neighbor, which is
three, and I paste a KNN classifier,

409
00:21:09,830 --> 00:21:11,670
and I train it.

410
00:21:11,670 --> 00:21:14,420
And after that I used-- this is
the prediction-- after I train it,

411
00:21:14,420 --> 00:21:17,930
I just used the model to
predict my testing data set,

412
00:21:17,930 --> 00:21:21,750
to see how accurate it is
on the data set itself.

413
00:21:21,750 --> 00:21:27,230
So, after I run it, you can see that
the accuracy score was around 93%.

414
00:21:27,230 --> 00:21:30,980
That means that of all the positive--

415
00:21:30,980 --> 00:21:36,480
of all of the predictions that all the
ML model made, 93% of them are correct.

416
00:21:36,480 --> 00:21:39,110
And what about the precision score?

417
00:21:39,110 --> 00:21:43,340
We got 0.9476, around 95%.

418
00:21:43,340 --> 00:21:46,040
That means that out of the
total positive observation,

419
00:21:46,040 --> 00:21:48,380
95% of the prediction
is actually correct.

420
00:21:48,380 --> 00:21:49,400
That's pretty high.

421
00:21:49,400 --> 00:21:52,190
And the recall of the KNN is around 93%.

422
00:21:52,190 --> 00:21:54,290
I'm so sorry, I will change this later.

423
00:21:54,290 --> 00:21:58,610
But, maybe, let's test around with
different kind of neighbor scores.

424
00:21:58,610 --> 00:22:03,600
[? Let's do ?] about 10, which
is basically overfeeding.

425
00:22:03,600 --> 00:22:06,600
We can see that it
actually affects the--

426
00:22:06,600 --> 00:22:09,270
I think, just now, we have 93% accuracy.

427
00:22:09,270 --> 00:22:13,500
But now, we can see that if we feed it
with a amount of 10 neighbors, which

428
00:22:13,500 --> 00:22:18,344
is increasing the radius of the
circle, you can see that the--

429
00:22:18,344 --> 00:22:20,680
both the precision, and
accuracy, and recall score

430
00:22:20,680 --> 00:22:24,190
actually increased
from around 93% to 95%.

431
00:22:24,190 --> 00:22:28,000
So that's why constantly
testing and evaluating the model

432
00:22:28,000 --> 00:22:29,540
is actually very important.

433
00:22:29,540 --> 00:22:34,670
So, let's move on to another one,
which is logistic regression.

434
00:22:34,670 --> 00:22:38,200
Logistic regression is
mainly based on this graph,

435
00:22:38,200 --> 00:22:44,060
called the MLE, maximum
likelihood estimation.

436
00:22:44,060 --> 00:22:48,340
Giving an example would be, say, for
example, I was given a bunch of data

437
00:22:48,340 --> 00:22:52,120
and I was trying to predict
whether someone is COVID positive

438
00:22:52,120 --> 00:22:53,630
or COVID negative.

439
00:22:53,630 --> 00:22:57,640
So, what we will do is, we will plot
a graph like this, a sigmoid curve.

440
00:22:57,640 --> 00:23:00,820
And then, say the data was
here, right, and I'll be like,

441
00:23:00,820 --> 00:23:04,900
oh, because this data was around more
than 90, more than half of this curve,

442
00:23:04,900 --> 00:23:06,880
I would classify it as one.

443
00:23:06,880 --> 00:23:10,310
But if it's here, we'll
classify it as COVID negative.

444
00:23:10,310 --> 00:23:13,030
So that's how this
model generally works,

445
00:23:13,030 --> 00:23:15,850
is that depending on
where the data points are,

446
00:23:15,850 --> 00:23:20,970
we classify whether it's positive
or negative based on where it is.

447
00:23:20,970 --> 00:23:24,050
So, in this case, you also try
to run a logistic regression

448
00:23:24,050 --> 00:23:26,760
model on the Iris data set itself.

449
00:23:26,760 --> 00:23:30,440
And I also include a few
resources where you can

450
00:23:30,440 --> 00:23:32,630
understand logistic regression more.

451
00:23:32,630 --> 00:23:41,962
So if we run the logistic
regressions I think it's actually--

452
00:23:41,962 --> 00:23:43,730
[INAUDIBLE]

453
00:23:43,730 --> 00:23:45,860
So, if you run the
logistic regression, you

454
00:23:45,860 --> 00:23:50,690
can see that the accuracy score
is actually less than the one

455
00:23:50,690 --> 00:23:51,710
that we had before.

456
00:23:51,710 --> 00:24:00,680
We have 95% of accuracy with KNN
model, but we only have around 93%

457
00:24:00,680 --> 00:24:02,870
with logistic regression.

458
00:24:02,870 --> 00:24:05,870
And that's how we actually, so--

459
00:24:05,870 --> 00:24:08,630
alternatively, you can
use other approach or ML

460
00:24:08,630 --> 00:24:10,400
model that we suggested here--

461
00:24:10,400 --> 00:24:11,930
sorry to keep jumping--

462
00:24:11,930 --> 00:24:15,290
which is decision tree, support
vector machine, or neural network.

463
00:24:15,290 --> 00:24:18,000
I remember when you run it
across support vector machine,

464
00:24:18,000 --> 00:24:20,840
it has really good
accuracy and prediction,

465
00:24:20,840 --> 00:24:24,750
while neural network is also a pretty
good way to try to classify it.

466
00:24:24,750 --> 00:24:27,740
So that's all for me you can run on to--

467
00:24:27,740 --> 00:24:33,560
KEVIN XU: Yeah, so, great,
Zad, a wonderful job

468
00:24:33,560 --> 00:24:38,550
of highlighting how you
can use a regression,

469
00:24:38,550 --> 00:24:46,062
or you can use ML to take a data set
and attempt to predict future elements

470
00:24:46,062 --> 00:24:47,770
that may be part of
this data set, right?

471
00:24:47,770 --> 00:24:50,580
And, so, this is one of
the core pieces of ML.

472
00:24:50,580 --> 00:24:54,220
And, don't be worried if you
couldn't follow all of that.

473
00:24:54,220 --> 00:24:58,950
Or, you don't really understand
how the syntax for the code works.

474
00:24:58,950 --> 00:25:01,740
It's learning new libraries,
and machine learning

475
00:25:01,740 --> 00:25:05,730
is very heavily dependent on
previously written libraries.

476
00:25:05,730 --> 00:25:11,500
It's a lot of work to develop your own
algorithm to-- for the machine learning

477
00:25:11,500 --> 00:25:12,390
take place.

478
00:25:12,390 --> 00:25:14,970
And, so, a lot of the case it's the--

479
00:25:14,970 --> 00:25:18,720
a lot of the time it's the case
of seeing how your data looks,

480
00:25:18,720 --> 00:25:21,150
and then trying to
choose the best model.

481
00:25:21,150 --> 00:25:24,510
Which is a sequence of algorithms
that make this black box function,

482
00:25:24,510 --> 00:25:29,050
as we talked about earlier, to
actually get accurate predictions.

483
00:25:29,050 --> 00:25:34,140
So, yeah, feel free to ask questions--
or pose questions about this.

484
00:25:34,140 --> 00:25:38,700
But I will be going on to kind of
the flip side of machine learning.

485
00:25:38,700 --> 00:25:44,640
Instead of using previous data to
predict elements of the same data set,

486
00:25:44,640 --> 00:25:50,670
we're going to start talking about how
can we get the machine to calculate

487
00:25:50,670 --> 00:25:53,820
or like-- we're going to start
talking about games, and reinforcement

488
00:25:53,820 --> 00:25:54,340
learning.

489
00:25:54,340 --> 00:25:57,870
And, so, one of the most popular
uses for reinforcement learning

490
00:25:57,870 --> 00:26:00,660
is, of course, trying
to solve games, right?

491
00:26:00,660 --> 00:26:06,780
So you have a game, and you want to
build an artificial intelligence,

492
00:26:06,780 --> 00:26:10,620
or an AI, that best
wins your game, or that

493
00:26:10,620 --> 00:26:14,470
always plays the best move possible.

494
00:26:14,470 --> 00:26:21,300
And this is actually loaded with theory
about how data sets work, how the game

495
00:26:21,300 --> 00:26:25,140
itself works, and there's a lot
of math and logic behind it.

496
00:26:25,140 --> 00:26:29,940
But, at the end of the day, the idea
is, given what you know about the game,

497
00:26:29,940 --> 00:26:33,300
can you get your computer
to train on the game such

498
00:26:33,300 --> 00:26:39,850
that it always has a pretty good
idea of what the next best move is?

499
00:26:39,850 --> 00:26:43,680
And, so, I've built just
a really silly game here.

500
00:26:43,680 --> 00:26:48,690
We don't have to worry too much about
the structure of the game itself,

501
00:26:48,690 --> 00:26:52,290
but in general, essentially,
when you start the game

502
00:26:52,290 --> 00:26:53,997
you're presented with two doors.

503
00:26:53,997 --> 00:26:55,830
You choose one of the
doors, and then you're

504
00:26:55,830 --> 00:26:57,330
presented with another two doors.

505
00:26:57,330 --> 00:27:01,020
And you choose one of the doors
again, and behind that door,

506
00:27:01,020 --> 00:27:05,970
there's a randomly
generated value between zero

507
00:27:05,970 --> 00:27:08,290
and a certain number that
corresponds with the door.

508
00:27:08,290 --> 00:27:11,730
So if you think about this
kind of tree-like structure,

509
00:27:11,730 --> 00:27:14,730
you have two, and then two more, so
there's four total doors at the end.

510
00:27:14,730 --> 00:27:16,855
And each one of them has
a number assigned to them.

511
00:27:16,855 --> 00:27:21,970
Perhaps like nine-- or in this case,
I think, three, nine, one, and 20.

512
00:27:21,970 --> 00:27:26,470
All right, so obviously, the
door that is associated with 20

513
00:27:26,470 --> 00:27:29,770
is going to, on average, give
you better points, or reward,

514
00:27:29,770 --> 00:27:33,880
or whatever this point
system works out as, right?

515
00:27:33,880 --> 00:27:36,440
Than the door that has a
one associated with it.

516
00:27:36,440 --> 00:27:39,310
And, so, the question is,
can you get the computer

517
00:27:39,310 --> 00:27:43,210
to realize which door
is the best, simply

518
00:27:43,210 --> 00:27:45,760
by playing the game a bunch of times?

519
00:27:45,760 --> 00:27:48,550
And this is the concept
of reinforcement learning.

520
00:27:48,550 --> 00:27:50,330
You just play it a bunch of times.

521
00:27:50,330 --> 00:27:54,220
And for things that turn out well,
as in you've got a lot of points,

522
00:27:54,220 --> 00:27:56,890
and well, yeah, you've
got a lot of points,

523
00:27:56,890 --> 00:28:00,790
then the computer should be more likely
to choose that option in the future.

524
00:28:00,790 --> 00:28:03,430
And for doors that you
didn't get a lot of points,

525
00:28:03,430 --> 00:28:06,890
you could be less likely to
choose that door in the future.

526
00:28:06,890 --> 00:28:11,890
So we don't have to worry too much about
the overall structure of this code,

527
00:28:11,890 --> 00:28:16,120
but right now, I have it set up
such that the computer just simply

528
00:28:16,120 --> 00:28:17,950
plays random moves every time.

529
00:28:17,950 --> 00:28:22,540
So it goes through one of the two
initial doors, with equal probability,

530
00:28:22,540 --> 00:28:25,730
and then it chooses one of the next
two doors with equal probability.

531
00:28:25,730 --> 00:28:29,080
So you expect that the
expected value is going

532
00:28:29,080 --> 00:28:32,950
to be pretty low, because these
aren't in general pretty--

533
00:28:32,950 --> 00:28:33,790
or, yeah.

534
00:28:33,790 --> 00:28:37,180
So it won't be the highest that
it could possibly be, right?

535
00:28:37,180 --> 00:28:45,880
So if we run this code, we see we
got an expected value about 4.125.

536
00:28:45,880 --> 00:28:50,843
So, on average, the computer
is scoring four points, right?

537
00:28:50,843 --> 00:28:53,260
And, so, this is definitely
not the best we can do, right?

538
00:28:53,260 --> 00:28:54,880
This is a totally random--

539
00:28:54,880 --> 00:28:56,620
the computer is playing
totally randomly,

540
00:28:56,620 --> 00:28:59,590
like this is completely
stupid for the computer to do.

541
00:28:59,590 --> 00:29:04,600
So we want to attempt to get it to
figure out which doors are better.

542
00:29:04,600 --> 00:29:06,790
And this actually doesn't have to--

543
00:29:06,790 --> 00:29:10,400
this-- to implement reinforcement
learning, a lot of the time neural

544
00:29:10,400 --> 00:29:12,400
networking and other
aspects of machine learning

545
00:29:12,400 --> 00:29:15,320
are incorporated with
reinforcement learning.

546
00:29:15,320 --> 00:29:21,280
So you take in the data, and you
pass it through another regression,

547
00:29:21,280 --> 00:29:25,790
or prediction, and that also helps
you find out which move is best.

548
00:29:25,790 --> 00:29:28,180
But, in this case, because
the game is so simple,

549
00:29:28,180 --> 00:29:31,740
you can simply just hard
code the training in.

550
00:29:31,740 --> 00:29:36,640
And, so, if we take a look here,
and this will be a little brief,

551
00:29:36,640 --> 00:29:39,140
so if you want to ask about the
logic, go ahead in the chat,

552
00:29:39,140 --> 00:29:41,600
but I'll try not to waste too
much time trying to explain

553
00:29:41,600 --> 00:29:43,950
what every line of code does.

554
00:29:43,950 --> 00:29:50,122
So for every move you take, you consider
the state that the move resides in.

555
00:29:50,122 --> 00:29:53,330
So, the current state of the board, so,
which doors you have in front of you,

556
00:29:53,330 --> 00:29:55,350
and which possible doors
you can go through.

557
00:29:55,350 --> 00:29:58,308
So, if you start at the beginning,
you have the first two doors, right?

558
00:29:58,308 --> 00:30:02,360
So, your state is just at the beginning,
and you have two possible choices.

559
00:30:02,360 --> 00:30:08,540
And, so, if you label every state in the
system with a heuristic value that is--

560
00:30:08,540 --> 00:30:11,690
that basically tells you
the goodness of that state,

561
00:30:11,690 --> 00:30:15,920
like how desirable do you are-- how
desirable is it to be in that state

562
00:30:15,920 --> 00:30:18,350
if the goal is to accumulate points.

563
00:30:18,350 --> 00:30:21,260
Then, what you can do is,
have the computer simply

564
00:30:21,260 --> 00:30:25,580
go through this, the states, that have
the highest goodness value, right?

565
00:30:25,580 --> 00:30:28,310
So how do we actually
calculate this goodness value?

566
00:30:28,310 --> 00:30:30,690
Well, we just play the
game a bunch of times.

567
00:30:30,690 --> 00:30:36,140
So for every time you play the game,
you're keeping an internal track,

568
00:30:36,140 --> 00:30:40,850
or the computer is keeping an internal
track, of what the current state

569
00:30:40,850 --> 00:30:42,350
heuristic value is.

570
00:30:42,350 --> 00:30:46,370
And then, it makes a move
based on what it thinks

571
00:30:46,370 --> 00:30:48,870
is the best move in this case.

572
00:30:48,870 --> 00:30:57,310
So, here, once we make the move, we
then find out what the next state is,

573
00:30:57,310 --> 00:31:00,580
and how good the next state
is as a result of the move.

574
00:31:00,580 --> 00:31:05,380
So using this accumulation, we
can kind of backtrack and see

575
00:31:05,380 --> 00:31:08,470
how good was the current state that
we were in before we made the move,

576
00:31:08,470 --> 00:31:10,610
and how good of a move did we make?

577
00:31:10,610 --> 00:31:13,330
And, so, of course, by
incrementing these, right, you

578
00:31:13,330 --> 00:31:17,740
increase the probability to make good
moves and you increase-- you decrease,

579
00:31:17,740 --> 00:31:20,270
relatively, the probability
to make bad moves.

580
00:31:20,270 --> 00:31:23,810


581
00:31:23,810 --> 00:31:33,790
And, so, if we throw this training in
that calculates the heuristic model,

582
00:31:33,790 --> 00:31:37,430
and we recompile, we see we got
an expected value of about 10,

583
00:31:37,430 --> 00:31:40,340
which is about twice as
big as we had previously.

584
00:31:40,340 --> 00:31:42,760
So the computer has gotten
way better at this game.

585
00:31:42,760 --> 00:31:45,462
And if you actually take a look
into the data structure, which

586
00:31:45,462 --> 00:31:47,920
I won't at the moment, because
it looks really complicated,

587
00:31:47,920 --> 00:31:50,620
and well, it's just a big
dictionary, and I don't think

588
00:31:50,620 --> 00:31:53,380
it will help you understand
what is going on here.

589
00:31:53,380 --> 00:31:56,020
If you actually take a look,
you'll see that the end result

590
00:31:56,020 --> 00:32:00,700
of the training, so the training,
we played to the game 10,000 times.

591
00:32:00,700 --> 00:32:07,810
And we evaluated the goodness of
every move for those 10,000 games.

592
00:32:07,810 --> 00:32:10,060
And you'll find that
after this training,

593
00:32:10,060 --> 00:32:16,150
the probability of choosing the
door that leads to this 20 door,

594
00:32:16,150 --> 00:32:20,350
sorry I went a little far, so that
leads to this 20 door at the end

595
00:32:20,350 --> 00:32:21,670
is almost one.

596
00:32:21,670 --> 00:32:24,170
It's like, 0.9997 or
something like that.

597
00:32:24,170 --> 00:32:27,460
And, so, the computer has
basically figured out,

598
00:32:27,460 --> 00:32:33,745
without us telling the computer anything
except for the game, right, the game

599
00:32:33,745 --> 00:32:40,070
state, and what the results
are, how to beat this game.

600
00:32:40,070 --> 00:32:43,550
And, so, this is the goal
of reinforcement learning.

601
00:32:43,550 --> 00:32:49,060
And I think this is a very
interesting thing, that has--

602
00:32:49,060 --> 00:32:52,300
there's a lot of application
in chess, and go,

603
00:32:52,300 --> 00:32:57,700
and this is the basic core idea of
how people have solved these games.

604
00:32:57,700 --> 00:33:02,830
You want to make good moves,
is what it boils down to.

605
00:33:02,830 --> 00:33:05,110
Which sounds very simple,
but in execution can

606
00:33:05,110 --> 00:33:09,190
be more complicated than it seems.

607
00:33:09,190 --> 00:33:09,690
Let's go--

608
00:33:09,690 --> 00:33:12,782
SPEAKER: Kevin, and Zad, can we
take a couple of questions now?

609
00:33:12,782 --> 00:33:13,740
Oh perfect timing, yay.

610
00:33:13,740 --> 00:33:16,140
KEVIN XU: --that's what we're
planning on doing right now.

611
00:33:16,140 --> 00:33:17,473
SPEAKER: There we go, all right.

612
00:33:17,473 --> 00:33:22,140
There have been several since
the last little break there.

613
00:33:22,140 --> 00:33:26,910
OK, from Angela, "Was Person
of Interest a realistic example

614
00:33:26,910 --> 00:33:29,940
of machine learning?"

615
00:33:29,940 --> 00:33:32,020
KEVIN XU: Person of Interest?

616
00:33:32,020 --> 00:33:34,660
ZAD CHIN: I'm sorry, can
you repeat the question?

617
00:33:34,660 --> 00:33:37,600
SPEAKER: That's the question, it
was probably about the earlier--

618
00:33:37,600 --> 00:33:40,090
maybe Angela, you can
write back in the chat?

619
00:33:40,090 --> 00:33:43,690
I'm going to just keep going
on and continue answering,

620
00:33:43,690 --> 00:33:46,720
but if you want to
clarify that in the chat.

621
00:33:46,720 --> 00:33:50,680
From Maria, "Do you ever feel
like machine learning sometimes

622
00:33:50,680 --> 00:33:56,400
have very serious sequences,
like in elections for example?"

623
00:33:56,400 --> 00:34:00,270
ZAD CHIN: Yes, we do feel like,
in terms of machine learning,

624
00:34:00,270 --> 00:34:02,760
the impact of machine
learning model, especially,

625
00:34:02,760 --> 00:34:04,470
I mean it can be both bad and good.

626
00:34:04,470 --> 00:34:06,730
That's why, at the last
slide of our slides,

627
00:34:06,730 --> 00:34:09,210
we also talk about how
machine learning actually

628
00:34:09,210 --> 00:34:15,635
generates deepfake, or like, privacy
intrusion, the alignment problem.

629
00:34:15,635 --> 00:34:17,760
So we actually include some
machine learning ethics

630
00:34:17,760 --> 00:34:20,080
that we hope to actually
share with you as well.

631
00:34:20,080 --> 00:34:25,179
But generally, in terms of
machine learning in an election,

632
00:34:25,179 --> 00:34:26,520
I do think that that's true.

633
00:34:26,520 --> 00:34:30,840
Because a lot of the time, advertisement
model in like Facebook, or Google,

634
00:34:30,840 --> 00:34:34,050
they use a lot of machine learning
model to predict the kind of person

635
00:34:34,050 --> 00:34:36,403
that you are, and recommend--

636
00:34:36,403 --> 00:34:38,820
recommender system is actually
a part of machine learning.

637
00:34:38,820 --> 00:34:41,440
It's a very huge research
topic in machine learning.

638
00:34:41,440 --> 00:34:46,199
And, so, if you want to know more it
will be recommender system at Google

639
00:34:46,199 --> 00:34:47,909
or Facebook, I would--

640
00:34:47,909 --> 00:34:52,440
SPEAKER: OK, and then we have
another question about cybersecurity.

641
00:34:52,440 --> 00:34:53,730
So is that--

642
00:34:53,730 --> 00:34:56,820
I think we can leave that for later,
I think you have some slides on that.

643
00:34:56,820 --> 00:34:58,530
Is that correct, Zad?

644
00:34:58,530 --> 00:34:59,390
Yeah, OK.

645
00:34:59,390 --> 00:35:00,528
ZAD CHIN: Yeah, we do.

646
00:35:00,528 --> 00:35:02,070
SPEAKER: All right, so, let me just--

647
00:35:02,070 --> 00:35:06,750
this is from James, "How does a
computer or machine recognize objects

648
00:35:06,750 --> 00:35:09,330
by itself in unsupervised learning?

649
00:35:09,330 --> 00:35:11,825


650
00:35:11,825 --> 00:35:12,700
ZAD CHIN: It depends.

651
00:35:12,700 --> 00:35:17,240
So for example, if you say like, a
data point, for example, currently

652
00:35:17,240 --> 00:35:19,610
I'm doing research on the ICU data set.

653
00:35:19,610 --> 00:35:24,650
We have a lot of features, and then, so
what we do is we paste in the features,

654
00:35:24,650 --> 00:35:29,840
and then it will be represented in
like, let me give you a simple example.

655
00:35:29,840 --> 00:35:32,630
Say, for example, we recommend--
we have x and y feature.

656
00:35:32,630 --> 00:35:34,190
And we recommend it on--

657
00:35:34,190 --> 00:35:37,238
we just plot, like we
just put the points in.

658
00:35:37,238 --> 00:35:40,530
And then how the machine learning knows
is that they try to cluster the points.

659
00:35:40,530 --> 00:35:45,110
One of the very good ways to actually
know, like in unsupervised learning,

660
00:35:45,110 --> 00:35:46,340
is clustering.

661
00:35:46,340 --> 00:35:48,620
So basically, say, for
example, I took a point, right?

662
00:35:48,620 --> 00:35:51,830
Like what is the neighbor of the
point, and how it should relate,

663
00:35:51,830 --> 00:35:53,260
how strong it actually relates.

664
00:35:53,260 --> 00:35:57,170
So, maybe like, for example,
the points are very separated,

665
00:35:57,170 --> 00:36:01,980
or the points are like a block, or
the points are not related at all.

666
00:36:01,980 --> 00:36:06,620
So, basically, one of the very
good ways of unsupervised learning

667
00:36:06,620 --> 00:36:10,310
is also clustering, that's
like for discrete data set.

668
00:36:10,310 --> 00:36:13,220
If you say for image data
set, for example, most

669
00:36:13,220 --> 00:36:16,880
of the time we use something called
a convolutional neural network, which

670
00:36:16,880 --> 00:36:19,840
is CNN, which is basically passed
through a lot of [? features. ?]

671
00:36:19,840 --> 00:36:23,540
And I think in CS50,
we also have this pset

672
00:36:23,540 --> 00:36:25,190
where we talk about edges detection.

673
00:36:25,190 --> 00:36:27,290
I think it's under the--

674
00:36:27,290 --> 00:36:28,610
it's one of the psets.

675
00:36:28,610 --> 00:36:33,500
And that's an edge detection,
whereby we draw out the edge itself.

676
00:36:33,500 --> 00:36:35,480
And that's also a part
of the CNN, whereby

677
00:36:35,480 --> 00:36:38,540
we try to pass through
different filters and networks.

678
00:36:38,540 --> 00:36:42,060
And then, we draw the edge to
recognize what is the image itself.

679
00:36:42,060 --> 00:36:46,880
So if you are interested in knowing
about how machine learning actually

680
00:36:46,880 --> 00:36:49,430
understands images, I
would recommend CNN.

681
00:36:49,430 --> 00:36:52,370
If you about-- if you like
normal, unsupervised learning,

682
00:36:52,370 --> 00:36:55,970
there are a few where we can say
here's like, clustering autoencoder,

683
00:36:55,970 --> 00:36:58,500
which is a part of neural network.

684
00:36:58,500 --> 00:37:01,710
SPEAKER: OK, why don't we take two
more, and then we'll continue again.

685
00:37:01,710 --> 00:37:04,800
So, Amir asks, "Which
step is preferred first?

686
00:37:04,800 --> 00:37:09,232
Data analysis or machine learning?"

687
00:37:09,232 --> 00:37:10,690
ZAD CHIN: I would-- yeah, go ahead.

688
00:37:10,690 --> 00:37:14,980
KEVIN XU: So, this is very
context-dependent at times.

689
00:37:14,980 --> 00:37:20,080
But in general, you don't want to take
some data set that you just gathered

690
00:37:20,080 --> 00:37:23,500
and immediately throw it
into machine learning.

691
00:37:23,500 --> 00:37:29,770
Because, when you take data, in the
real world, as opposed to generate data,

692
00:37:29,770 --> 00:37:31,250
there's a lot of noise.

693
00:37:31,250 --> 00:37:33,970
There's a lot of
inconsistencies, there is

694
00:37:33,970 --> 00:37:37,100
a lot of things that can go wrong,
just when you take normal data.

695
00:37:37,100 --> 00:37:42,040
And, so, if you just throw
something random into an ML program,

696
00:37:42,040 --> 00:37:45,140
it will not necessarily get
you the results you want.

697
00:37:45,140 --> 00:37:48,910
And most of the time it won't, because
there's so much noise in the data

698
00:37:48,910 --> 00:37:52,960
that it's really difficult
to identify the patterns.

699
00:37:52,960 --> 00:37:57,820
And, so, generally, when
you're designing this ML--

700
00:37:57,820 --> 00:38:04,060
type of sequence of steps, you want to
at least screen your data before you

701
00:38:04,060 --> 00:38:06,010
throw it into any type of program.

702
00:38:06,010 --> 00:38:10,300
That way, you can catch early on if
things are going to go wrong, right?

703
00:38:10,300 --> 00:38:12,550
And like, in the worst
case, you just might

704
00:38:12,550 --> 00:38:16,410
have to retake the entire data set
because it's not valuable, right?

705
00:38:16,410 --> 00:38:20,230
And, so, it's kind of like a screening
process, at least the initial data

706
00:38:20,230 --> 00:38:24,358
analysis, before you can actually try
and be productive with that data set.

707
00:38:24,358 --> 00:38:26,650
ZAD CHIN: Adding it on, I
feel like it's very important

708
00:38:26,650 --> 00:38:27,832
to do data analysis first.

709
00:38:27,832 --> 00:38:30,790
This is because say, for example,
you get a data of like COVID positive

710
00:38:30,790 --> 00:38:32,320
and COVID negative patients.

711
00:38:32,320 --> 00:38:34,512
And actually it is very dangerous.

712
00:38:34,512 --> 00:38:36,220
So, for example, if
your data set is very

713
00:38:36,220 --> 00:38:38,950
biased toward COVID negative
person, and you just

714
00:38:38,950 --> 00:38:41,800
pass it to a logistic regression
model, and the model would

715
00:38:41,800 --> 00:38:45,670
be like, oh, since 98% of the
people are actually COVID negative,

716
00:38:45,670 --> 00:38:48,790
right, then I can just
predict, oh nine--

717
00:38:48,790 --> 00:38:53,050
out of the test example I gave, I will
just predict everybody as negative.

718
00:38:53,050 --> 00:38:56,920
So I still get a 98% accuracy, which
is actually very, very dangerous.

719
00:38:56,920 --> 00:38:57,940
And it's all the way--

720
00:38:57,940 --> 00:39:00,995
why you want to know how
many labels that we have,

721
00:39:00,995 --> 00:39:03,370
before we actually pass it to
the machine learning model.

722
00:39:03,370 --> 00:39:05,350
Because these kind of biases can happen.

723
00:39:05,350 --> 00:39:10,578
Machines can be just like, oh, because
98% of people are actually negative,

724
00:39:10,578 --> 00:39:12,370
so I can just predict
everyone is negative.

725
00:39:12,370 --> 00:39:14,980
And I get a 98% accuracy, right?

726
00:39:14,980 --> 00:39:18,040
So it's very, very important
to do the data analysis

727
00:39:18,040 --> 00:39:20,320
before you actually train
the data, especially

728
00:39:20,320 --> 00:39:24,040
on this kind of highly
biased data itself.

729
00:39:24,040 --> 00:39:26,830
SPEAKER: OK, and the last question,
from Victor, "Do all machine

730
00:39:26,830 --> 00:39:28,870
learning models use neutral--

731
00:39:28,870 --> 00:39:32,438
neural networks?"

732
00:39:32,438 --> 00:39:32,980
ZAD CHIN: No.

733
00:39:32,980 --> 00:39:33,772
KEVIN XU: Yeah, no.

734
00:39:33,772 --> 00:39:35,770
So neural networking are--

735
00:39:35,770 --> 00:39:38,440
convolutional neural
networking, these two things,

736
00:39:38,440 --> 00:39:43,540
or same, thing but specialized,
are a specific subset

737
00:39:43,540 --> 00:39:45,680
of the machine learning that you can do.

738
00:39:45,680 --> 00:39:48,430
And, so, I think you probably
asked this before I showed

739
00:39:48,430 --> 00:39:50,200
the reinforcement learning example.

740
00:39:50,200 --> 00:39:53,440
But you don't have to use neural
networks at all when you're

741
00:39:53,440 --> 00:39:57,210
trying to get the machine to learn.

742
00:39:57,210 --> 00:40:02,400
It is very-- neural networks are
very powerful, because you're

743
00:40:02,400 --> 00:40:07,680
able to take a lot of data
and have the computer generate

744
00:40:07,680 --> 00:40:09,720
the relationships between the data.

745
00:40:09,720 --> 00:40:14,610
But it's not always necessary
to use a neural network when

746
00:40:14,610 --> 00:40:18,390
you're trying to learn from a data set,
or have the computer learn from a data

747
00:40:18,390 --> 00:40:20,940
set, as you saw in
reinforcement learning, right?

748
00:40:20,940 --> 00:40:26,430
You can simply have the computer attempt
to make its own conclusions based

749
00:40:26,430 --> 00:40:32,640
on what state, or what type of data you
give it, and what it wants to achieve.

750
00:40:32,640 --> 00:40:36,320


751
00:40:36,320 --> 00:40:37,280
ZAD CHIN: Right.

752
00:40:37,280 --> 00:40:41,520
KEVIN XU: So, yeah, we'll take
some more questions at the end.

753
00:40:41,520 --> 00:40:44,720
So yeah, feel free to continue
posting them in the chat.

754
00:40:44,720 --> 00:40:50,150
But, for now, since we've probably hit
you with a lot, and things we can do,

755
00:40:50,150 --> 00:40:52,550
and things we can't do,
and possible things,

756
00:40:52,550 --> 00:40:56,120
we want to give some tips on
what is reasonable to consider

757
00:40:56,120 --> 00:41:00,300
if you're actually attempting this-- to
implement this in a CS50 final project.

758
00:41:00,300 --> 00:41:03,980
And, so, Zad has made
this great graph here

759
00:41:03,980 --> 00:41:06,710
that goes kind of in difficulty
level from the left to right.

760
00:41:06,710 --> 00:41:09,260
And, so, at the very left
you have supervised learning

761
00:41:09,260 --> 00:41:12,260
and unsupervised learning, which is--

762
00:41:12,260 --> 00:41:16,880
requires some effort, but not
huge amounts of dedication.

763
00:41:16,880 --> 00:41:19,170
Although, this is always
context-dependent as well.

764
00:41:19,170 --> 00:41:23,480
And then, reinforcement learning
will probably take more time, simply

765
00:41:23,480 --> 00:41:26,600
because you not only
have to provide the data,

766
00:41:26,600 --> 00:41:29,180
but you often have to build
an infrastructure that

767
00:41:29,180 --> 00:41:30,420
can interpret the data.

768
00:41:30,420 --> 00:41:35,540
So, in the case of the game, right, you
have to build something that actually--

769
00:41:35,540 --> 00:41:38,090
you have to actually build
the game into Python.

770
00:41:38,090 --> 00:41:42,170
Where the game has to take
in an input state somehow,

771
00:41:42,170 --> 00:41:45,770
and it has to return to you
the new state and the results.

772
00:41:45,770 --> 00:41:49,010
And, so, there's all
this extra infrastructure

773
00:41:49,010 --> 00:41:52,040
that you need before you
can actually run any ML,

774
00:41:52,040 --> 00:41:55,760
and sometimes this takes longer
than running the actual ML.

775
00:41:55,760 --> 00:41:58,190
I actually spent longer trying
to get this infrastructure

776
00:41:58,190 --> 00:42:01,950
to work out than actually implementing
the reinforcement learning.

777
00:42:01,950 --> 00:42:04,160
So this is like very--

778
00:42:04,160 --> 00:42:07,160
something to be cautious of, if
you're interested in doing something

779
00:42:07,160 --> 00:42:09,440
like solving a game.

780
00:42:09,440 --> 00:42:13,550
But, of course, it is
certainly doable if you

781
00:42:13,550 --> 00:42:16,800
are willing to put in
the extra time to do it.

782
00:42:16,800 --> 00:42:20,090
And then, with convolutional
neural networks, and deep learning,

783
00:42:20,090 --> 00:42:24,500
and some of the higher
stuff, we caution against it

784
00:42:24,500 --> 00:42:27,530
unless you are very familiar
with this kind of construct.

785
00:42:27,530 --> 00:42:32,000
And it requires some-- quite
a bit of in depth knowledge.

786
00:42:32,000 --> 00:42:35,700
So you don't really have
to worry about that.

787
00:42:35,700 --> 00:42:39,290
And, so, quickly, I just want
to mention how you can actually

788
00:42:39,290 --> 00:42:40,430
implement these things.

789
00:42:40,430 --> 00:42:45,710
So we used Google Colab which
runs Jupyter Notebook which

790
00:42:45,710 --> 00:42:48,630
is a Python interpreter
that goes line by line.

791
00:42:48,630 --> 00:42:52,500
So the nice thing about Google Colab
is that, well, there's two nice things.

792
00:42:52,500 --> 00:42:54,570
One, is that it interprets
the line by line.

793
00:42:54,570 --> 00:42:57,320
So you can change a line
without having to rerun

794
00:42:57,320 --> 00:42:59,660
the script from the top
down, which is great

795
00:42:59,660 --> 00:43:04,220
if your things take forever to run, as
is often the case in machine learning.

796
00:43:04,220 --> 00:43:08,490
And the other nice thing about Google
Colab is that it's cloud computed.

797
00:43:08,490 --> 00:43:11,540
So there's a GPU on the
server end that does all this,

798
00:43:11,540 --> 00:43:14,670
and then returns it to you over the web.

799
00:43:14,670 --> 00:43:17,120
And, so you won't try
and break your machine

800
00:43:17,120 --> 00:43:21,380
trying to process like 10,000 images.

801
00:43:21,380 --> 00:43:25,010
But, on the other hand, if you feel
comfortable running it on your device,

802
00:43:25,010 --> 00:43:26,900
it's definitely certainly doable.

803
00:43:26,900 --> 00:43:29,720
I ran the reinforcement
learning code just fine

804
00:43:29,720 --> 00:43:33,120
on just a terminal on my device.

805
00:43:33,120 --> 00:43:35,570
And that-- those are the kind
of things that don't really

806
00:43:35,570 --> 00:43:37,220
take too much processing.

807
00:43:37,220 --> 00:43:42,230
So, yeah, just keep that in
mind when you are thinking

808
00:43:42,230 --> 00:43:46,840
about how to actually implement this.

809
00:43:46,840 --> 00:43:49,530
ZAD CHIN: So, next, I will talk
about the useful Python library

810
00:43:49,530 --> 00:43:50,790
that we almost--

811
00:43:50,790 --> 00:43:53,380
not-- normally use for machine learning.

812
00:43:53,380 --> 00:43:55,360
The first one we will go into is data.

813
00:43:55,360 --> 00:43:56,760
How do we get data, right?

814
00:43:56,760 --> 00:43:59,610
So, in terms of mining
data from online, we

815
00:43:59,610 --> 00:44:03,458
can go for BeautifulSoup, which
is kind of like a scrap library

816
00:44:03,458 --> 00:44:04,500
that someone wrote about.

817
00:44:04,500 --> 00:44:06,390
It's all linked, so you
guys can press that,

818
00:44:06,390 --> 00:44:09,450
and the slides are available
on the CS50 website.

819
00:44:09,450 --> 00:44:13,860
And, also, Scrapy, which is
a very good scraping library.

820
00:44:13,860 --> 00:44:18,550
And in some of like, built in data
sources that are very well documented,

821
00:44:18,550 --> 00:44:21,360
I would recommend Kaggle,
it's a Google platform

822
00:44:21,360 --> 00:44:23,100
with a lot of machine learning data.

823
00:44:23,100 --> 00:44:27,480
And UCI, University of California
Irvine machine learning repo--

824
00:44:27,480 --> 00:44:29,920
also a lot of data sets available.

825
00:44:29,920 --> 00:44:32,490
And you can also scrap
from websites or API calls,

826
00:44:32,490 --> 00:44:35,910
with BeautifulSoup or Scrapy,
or even your own API call.

827
00:44:35,910 --> 00:44:39,390
And, for data pre-processing,
I'm a very huge fan of Pandas.

828
00:44:39,390 --> 00:44:42,090
So I highly, highly
recommend using Pandas,

829
00:44:42,090 --> 00:44:45,390
like changing your CSV to
Pandas, and just work from there.

830
00:44:45,390 --> 00:44:47,560
It's actually much more
efficient and useful.

831
00:44:47,560 --> 00:44:51,540
And there's also NumPy and SciPy, which
is also like things that you normally

832
00:44:51,540 --> 00:44:54,040
use in terms of
calculating mean, median,

833
00:44:54,040 --> 00:44:56,580
and a lot of very useful
functions to analyze

834
00:44:56,580 --> 00:44:59,850
the data or pre-process the data.

835
00:44:59,850 --> 00:45:03,920
So the next thing is Python library,
which is like visualization.

836
00:45:03,920 --> 00:45:08,150
The most simple basic visualization
library that is available online is

837
00:45:08,150 --> 00:45:11,030
Matplotlib, very useful,
super well documented,

838
00:45:11,030 --> 00:45:13,370
a lot of examples
online that you can use.

839
00:45:13,370 --> 00:45:16,100
Seaborn is a better
kind of visualization,

840
00:45:16,100 --> 00:45:19,760
where you can choose your own color
map, or better kind of visualization.

841
00:45:19,760 --> 00:45:23,000
The best visualization that I can
actually think of, maybe not the best,

842
00:45:23,000 --> 00:45:23,930
but like--

843
00:45:23,930 --> 00:45:26,030
Plotly is a very highly visual--

844
00:45:26,030 --> 00:45:31,460
very engaging, highly visual,
very nice visualization libraries

845
00:45:31,460 --> 00:45:33,510
that Python have.

846
00:45:33,510 --> 00:45:37,220
So, in terms of ML models, there
are a lot of built in libraries.

847
00:45:37,220 --> 00:45:39,590
You can also do your own
ML model from scratch,

848
00:45:39,590 --> 00:45:44,190
from Python, which actually increases
your understanding on the model itself.

849
00:45:44,190 --> 00:45:47,840
But if you want to save time, if you
just want to use the model itself,

850
00:45:47,840 --> 00:45:49,490
there are a lot of built in libraries.

851
00:45:49,490 --> 00:45:52,820
Say, for example, SKlearn is a
very good library for beginners,

852
00:45:52,820 --> 00:45:56,240
they are various classification,
regression, and clustering algorithm

853
00:45:56,240 --> 00:45:57,620
that we can just use.

854
00:45:57,620 --> 00:46:02,110
And there is also TensorFlow,
I think a Google kind of like--

855
00:46:02,110 --> 00:46:05,650
a Google deep neural
network, Python library.

856
00:46:05,650 --> 00:46:08,350
It goes for various tasks,
it's on training and inference,

857
00:46:08,350 --> 00:46:09,700
mostly on deep neural network.

858
00:46:09,700 --> 00:46:13,690
And they have pretty good documentation
online and on YouTube and really

859
00:46:13,690 --> 00:46:15,190
good tutorials.

860
00:46:15,190 --> 00:46:18,640
We also have PyTorch, where it also
runs on neural network computer

861
00:46:18,640 --> 00:46:20,410
vision and NLP.

862
00:46:20,410 --> 00:46:23,140
And we have Keras, I think
Keras is by Facebook.

863
00:46:23,140 --> 00:46:26,140
And it's primarily for developing
and evaluating deep learning

864
00:46:26,140 --> 00:46:26,765
models as well.

865
00:46:26,765 --> 00:46:28,932
Those are really, really
useful libraries that you--

866
00:46:28,932 --> 00:46:31,480
and are really well documented,
there are a lot of tutorials

867
00:46:31,480 --> 00:46:33,970
online that we actually
recommend, and it also

868
00:46:33,970 --> 00:46:36,700
linked so you guys can
actually have a look too.

869
00:46:36,700 --> 00:46:40,270
In terms of ML resources and
support from CS50 and beyond,

870
00:46:40,270 --> 00:46:43,330
we actually recommend you to
go on Ed if you have any bugs

871
00:46:43,330 --> 00:46:44,920
that you don't know how to fix.

872
00:46:44,920 --> 00:46:48,410
The team-- the CS50 team is really,
really happy to help you on Ed.

873
00:46:48,410 --> 00:46:51,460
And you also have CS50
Intro to AI classes

874
00:46:51,460 --> 00:46:53,830
online too, so you guys
can have a look at it.

875
00:46:53,830 --> 00:46:58,040
They also a lot of Python libraries
documentation online, where I highly,

876
00:46:58,040 --> 00:47:00,400
highly recommend you to
look over the examples

877
00:47:00,400 --> 00:47:02,270
before you actually start coding.

878
00:47:02,270 --> 00:47:05,470
Kaggle is a place where
all the data scientists go.

879
00:47:05,470 --> 00:47:10,810
They are data sets, they are also
example codes that you can try out,

880
00:47:10,810 --> 00:47:12,820
and also competitions that you can join.

881
00:47:12,820 --> 00:47:15,160
Very good place to learn
about data science, and more

882
00:47:15,160 --> 00:47:16,580
about machine learning.

883
00:47:16,580 --> 00:47:19,570
And there are a few blogs that
we found very useful as tools.

884
00:47:19,570 --> 00:47:22,675
Data science blog is by Medium,
and Machine Learning Mastery blog

885
00:47:22,675 --> 00:47:24,550
is a free blog that's
really, really helpful.

886
00:47:24,550 --> 00:47:27,920
Of course, we have our favorite,
Stack Overflow and GitHub, really,

887
00:47:27,920 --> 00:47:31,400
really good ML resources
and support there as well.

888
00:47:31,400 --> 00:47:34,390
So let's talk about the
ML ethics that we actually

889
00:47:34,390 --> 00:47:37,390
see, like a lot of questions
about what is the best.

890
00:47:37,390 --> 00:47:39,700
Machine learning can do
so much good stuff, right?

891
00:47:39,700 --> 00:47:42,263
Like it is used in health
care, education, anywhere

892
00:47:42,263 --> 00:47:43,180
that you can think of.

893
00:47:43,180 --> 00:47:47,450
But, at the same time, it's had
its own like, danger itself.

894
00:47:47,450 --> 00:47:50,650
So, for example, one of the things
that we see the most is deepfake.

895
00:47:50,650 --> 00:47:54,670
And I think Malan actually did a
deepfake example video in last year's

896
00:47:54,670 --> 00:47:56,290
CS50, which is actually very exciting.

897
00:47:56,290 --> 00:47:58,480
I highly, highly recommend
you guys to watch,

898
00:47:58,480 --> 00:48:01,330
I actually linked it here so
you guys can watch it later.

899
00:48:01,330 --> 00:48:05,480
And another thing about machine
learning is actually it's a black box.

900
00:48:05,480 --> 00:48:08,080
So it's like, interpretability
in machine learning

901
00:48:08,080 --> 00:48:11,530
is a huge topic that a lot of
machine learning practitioners

902
00:48:11,530 --> 00:48:12,650
actually talking about.

903
00:48:12,650 --> 00:48:16,480
So one of them is Professor Finale,
also a professor at Harvard.

904
00:48:16,480 --> 00:48:20,500
He-- she actually does a lot of
stuff about interpretability in AI

905
00:48:20,500 --> 00:48:22,010
and in healthcare as well.

906
00:48:22,010 --> 00:48:25,350
So I highly recommend you to watch
the TED Talk with her, if you want to.

907
00:48:25,350 --> 00:48:27,100
And the other thing
about machine learning

908
00:48:27,100 --> 00:48:29,660
is AI biases, fairness,
we heard a lot about it.

909
00:48:29,660 --> 00:48:32,560
So there is this specific
course by MIT that talks

910
00:48:32,560 --> 00:48:35,590
about AI biases and fairness,
really well documented video,

911
00:48:35,590 --> 00:48:37,030
highly recommend.

912
00:48:37,030 --> 00:48:41,832
And like, how you guys say we have
a lot of fairness, transparency,

913
00:48:41,832 --> 00:48:44,040
privacy issues, that are
related to machine learning.

914
00:48:44,040 --> 00:48:46,990
And there's this book that is really
good that I wrote here as well.

915
00:48:46,990 --> 00:48:49,590
It's called The Alignment Problem,
it's by a professor in UCB.

916
00:48:49,590 --> 00:48:52,620
It's about how we can align machine
learning with our human values.

917
00:48:52,620 --> 00:48:55,770
All of this stuff is
linked, so you have more--

918
00:48:55,770 --> 00:48:59,070
if you like to know more
about machine learning ethics

919
00:48:59,070 --> 00:49:01,440
and how it actually can be dangerous.

920
00:49:01,440 --> 00:49:03,750
But I don't say I don't
support it, I was like,

921
00:49:03,750 --> 00:49:06,210
machine learning is very helpful,
but we need to be mindful

922
00:49:06,210 --> 00:49:09,580
that it can be perilous as well.

923
00:49:09,580 --> 00:49:13,770
So, that's all from us, and
we will take some questions.

924
00:49:13,770 --> 00:49:15,330
And, yeah.

925
00:49:15,330 --> 00:49:16,830
SPEAKER: OK, wonderful.

926
00:49:16,830 --> 00:49:20,970
So, let's go to Doris, "Is there a
need for bigger capacity of laptop

927
00:49:20,970 --> 00:49:21,960
for machine learning?

928
00:49:21,960 --> 00:49:24,750
I'm using an old Mac."

929
00:49:24,750 --> 00:49:27,400
KEVIN XU: This is
actually a great question.

930
00:49:27,400 --> 00:49:32,280
This is truly dependent on the
data sets that you're working with.

931
00:49:32,280 --> 00:49:34,860
A lot of the times, at
least in the universities,

932
00:49:34,860 --> 00:49:38,160
right, when they're dealing
with huge amounts of data,

933
00:49:38,160 --> 00:49:40,150
they have to process
this on the cluster.

934
00:49:40,150 --> 00:49:43,500
Which is a cluster of
server computers that will--

935
00:49:43,500 --> 00:49:46,680
that are like way higher
tech than anything

936
00:49:46,680 --> 00:49:48,300
you could purchase individually.

937
00:49:48,300 --> 00:49:51,210
But, for the sake of
implementing a small ML

938
00:49:51,210 --> 00:49:53,550
project, for the sake
of learning about ML,

939
00:49:53,550 --> 00:49:58,560
or for the sake of doing a fun small
project, such as like attempting

940
00:49:58,560 --> 00:50:01,750
to write an image pro--
like, recognitions

941
00:50:01,750 --> 00:50:04,200
set, with just a bunch of
images, this is something

942
00:50:04,200 --> 00:50:08,730
that is very doable on most machines.

943
00:50:08,730 --> 00:50:12,456
Of course, it really depends on like--

944
00:50:12,456 --> 00:50:15,960
yeah, it will be software--
or hardware dependent here,

945
00:50:15,960 --> 00:50:23,360
but I would say that short of
being very, very old hardware,

946
00:50:23,360 --> 00:50:26,510
then you should be able to at
least get the script working.

947
00:50:26,510 --> 00:50:30,950
But, of course, if it doesn't,
cloud computing is always an option.

948
00:50:30,950 --> 00:50:34,370
Google Colab is free,
which is great, so it's--

949
00:50:34,370 --> 00:50:38,880
and it's honestly not any different
from just accessing the [? IDE. ?]

950
00:50:38,880 --> 00:50:42,200
And, so, we really
recommend that if you are

951
00:50:42,200 --> 00:50:45,740
interested in things that will
run really slowly, that you

952
00:50:45,740 --> 00:50:48,600
look into cloud on Google Colab.

953
00:50:48,600 --> 00:50:52,410
SPEAKER: OK, [? Aviral, ?]
"Is it possible

954
00:50:52,410 --> 00:50:56,700
to use two types of different
algorithms at the same time

955
00:50:56,700 --> 00:51:01,330
to increase the accuracy of the model?"

956
00:51:01,330 --> 00:51:04,320
ZAD CHIN: It's not two possible
models at the same time.

957
00:51:04,320 --> 00:51:05,775
I think it is--

958
00:51:05,775 --> 00:51:08,310
normally like we actually
use different model,

959
00:51:08,310 --> 00:51:11,250
and different model has its own
strengths and weaknesses, I would say.

960
00:51:11,250 --> 00:51:14,980
So, for example, like KNN,
it might overfit very well.

961
00:51:14,980 --> 00:51:18,570
But, then, I think, for example,
like mo-- the example that you

962
00:51:18,570 --> 00:51:21,120
have, KNN versus
logistic regression, it's

963
00:51:21,120 --> 00:51:23,290
not like we can integrate
both two together.

964
00:51:23,290 --> 00:51:27,450
I think both KNN and logistic
regression has its own advantages

965
00:51:27,450 --> 00:51:28,600
and disadvantages.

966
00:51:28,600 --> 00:51:33,980
The idea is to test it against
different kinds of ML models

967
00:51:33,980 --> 00:51:35,620
that's available out there.

968
00:51:35,620 --> 00:51:40,510
And to increase-- to see which in ML
performs better on the current data

969
00:51:40,510 --> 00:51:41,010
set.

970
00:51:41,010 --> 00:51:43,590
So for example you can see
that KNN performed really

971
00:51:43,590 --> 00:51:45,450
well on the data set
that you have just now,

972
00:51:45,450 --> 00:51:48,927
but then it doesn't mean that KNN will
always perform better on other data

973
00:51:48,927 --> 00:51:50,010
sets that you are testing.

974
00:51:50,010 --> 00:51:52,552
So it's actually very recommended
to use different data sets.

975
00:51:52,552 --> 00:51:56,190
But on the idea of integrating two
machine learning models together

976
00:51:56,190 --> 00:51:59,400
to get higher accuracy,
it might be possible,

977
00:51:59,400 --> 00:52:01,750
but I'm not really sure about it.

978
00:52:01,750 --> 00:52:05,850
KEVIN XU: I would say the close analogy
is that if you can recall-- well

979
00:52:05,850 --> 00:52:07,470
actually, let's just go back to it.

980
00:52:07,470 --> 00:52:11,160
If you can recall the
nodes that we have, right?

981
00:52:11,160 --> 00:52:14,790
You can think of each set of nodes
as like, a different transformation

982
00:52:14,790 --> 00:52:15,970
that you do to the data.

983
00:52:15,970 --> 00:52:20,980
So what is often the case is that your
data will go through many, many layers,

984
00:52:20,980 --> 00:52:23,370
especially for things
that are very complicated.

985
00:52:23,370 --> 00:52:26,370
And, so, this is not necessarily
applying two different machine learning

986
00:52:26,370 --> 00:52:28,740
models, but you are--

987
00:52:28,740 --> 00:52:31,210
this kind of transformation
is very iterative.

988
00:52:31,210 --> 00:52:34,590
And, so, this is like--

989
00:52:34,590 --> 00:52:37,950
we call them layers, so like you
pass the data through one layer,

990
00:52:37,950 --> 00:52:40,980
and then you pass it through another
layer, iteratively, until you get

991
00:52:40,980 --> 00:52:42,970
to this end output layer.

992
00:52:42,970 --> 00:52:46,440
So this actually goes into the
specific of how the function works,

993
00:52:46,440 --> 00:52:50,550
but you will often have
to make multiple layers

994
00:52:50,550 --> 00:52:56,250
to get good results for your data set,
because the data is always complicated.

995
00:52:56,250 --> 00:52:57,300
SPEAKER: This is, an I--

996
00:52:57,300 --> 00:53:00,450
please excuse my pronunciation
here, [? Effie ?]

997
00:53:00,450 --> 00:53:03,330
writes, "Does ML help in cybersecurity?"

998
00:53:03,330 --> 00:53:07,330


999
00:53:07,330 --> 00:53:09,997
ZAD CHIN: Wait, the question is,
"Does ML affect cyber security?

1000
00:53:09,997 --> 00:53:10,913
KEVIN XU: Or does it--

1001
00:53:10,913 --> 00:53:12,810
SPEAKER: "Does ML help
in cybersecurity?"

1002
00:53:12,810 --> 00:53:14,780
ZAD CHIN: Oh, I think it does.

1003
00:53:14,780 --> 00:53:17,560
So, for example, I think I
heard a lot of information

1004
00:53:17,560 --> 00:53:22,070
about how banks actually use ML
to detect fraud transactions.

1005
00:53:22,070 --> 00:53:24,250
So in terms of that kind
of way, there's a lot

1006
00:53:24,250 --> 00:53:29,350
of ways whereby ML actually helps in
trying to prevent cybersecurity attacks

1007
00:53:29,350 --> 00:53:33,040
or like detect cybersecurity attacks,
especially in a large company.

1008
00:53:33,040 --> 00:53:35,290
Because it's actually
really well in trying

1009
00:53:35,290 --> 00:53:39,370
to find patterns, or unusual patterns
in a huge amount of data set.

1010
00:53:39,370 --> 00:53:43,690
So I definitely think that it has a
very huge application in cybersecurity

1011
00:53:43,690 --> 00:53:44,688
itself.

1012
00:53:44,688 --> 00:53:46,480
KEVIN XU: Yeah, and if
there's one takeaway

1013
00:53:46,480 --> 00:53:51,010
that we want you to have from
this, it's that the goal of-- what

1014
00:53:51,010 --> 00:53:53,650
ML is really good at doing
is taking a lot of data

1015
00:53:53,650 --> 00:53:56,360
and finding connections
between those data points.

1016
00:53:56,360 --> 00:54:00,820
So anything you can think of that
needs to process a huge amount of data,

1017
00:54:00,820 --> 00:54:04,120
you can almost always
apply some form of ML.

1018
00:54:04,120 --> 00:54:08,210
If this data is meaningful, of course.

1019
00:54:08,210 --> 00:54:11,710
So in terms of
cybersecurity, like actually

1020
00:54:11,710 --> 00:54:14,870
building firewalls and
those kinds of things,

1021
00:54:14,870 --> 00:54:16,960
it doesn't have to
necessarily do as much

1022
00:54:16,960 --> 00:54:19,120
with processing huge amounts of data.

1023
00:54:19,120 --> 00:54:23,108
But, what, as Zad said, right,
with like, fraudulent account

1024
00:54:23,108 --> 00:54:25,150
accesses, and things like
that, that is something

1025
00:54:25,150 --> 00:54:28,500
that is totally very applicable to ML.

1026
00:54:28,500 --> 00:54:29,610
SPEAKER: OK.

1027
00:54:29,610 --> 00:54:34,200
[? Madhi ?] asks, "What is the AI
application in video games nowadays?

1028
00:54:34,200 --> 00:54:40,040
And as a machine learning
developer, can I work in game field?

1029
00:54:40,040 --> 00:54:42,400
KEVIN XU: Oh yeah, so, I
can take this question.

1030
00:54:42,400 --> 00:54:47,880
Yeah, so, I don't know if you've seen,
but the I think the team, OpenAI,

1031
00:54:47,880 --> 00:54:52,710
I think is run by Tesla, or is
like a subsidiary or maybe--

1032
00:54:52,710 --> 00:54:53,960
ZAD CHIN: I think it's Google.

1033
00:54:53,960 --> 00:54:55,418
KEVIN XU: It might be Google, yeah.

1034
00:54:55,418 --> 00:54:59,040
So OpenAI is a team that has
been developing machine learning

1035
00:54:59,040 --> 00:55:01,290
game software for quite a while.

1036
00:55:01,290 --> 00:55:06,420
And a couple of years ago, they
released, like, big news, Dota 2, which

1037
00:55:06,420 --> 00:55:12,120
is one of the popular
MOBA games, they managed

1038
00:55:12,120 --> 00:55:14,720
to write an AI that beat
a team of professionals.

1039
00:55:14,720 --> 00:55:18,300
So this is totally
something that is doable,

1040
00:55:18,300 --> 00:55:25,853
and currently, and I recommend
you check out the OpenAI--

1041
00:55:25,853 --> 00:55:27,270
the stuff that they've been doing.

1042
00:55:27,270 --> 00:55:32,258
It's like, really cool and
it's definitely very marketable

1043
00:55:32,258 --> 00:55:34,050
if you're interested
in that kind of stuff.

1044
00:55:34,050 --> 00:55:37,183
ZAD CHIN: I think in terms of games,
like DeepMind also created AlphaGo,

1045
00:55:37,183 --> 00:55:38,100
which is like one of--

1046
00:55:38,100 --> 00:55:41,310
I don't know whether that's considered
a video game, but it's a game.

1047
00:55:41,310 --> 00:55:45,150
It's just super huge at that time,
because it's the world's best Go

1048
00:55:45,150 --> 00:55:46,500
player on planet Earth.

1049
00:55:46,500 --> 00:55:48,180
It was like AlphaGo versus human.

1050
00:55:48,180 --> 00:55:51,960
There's also a movie about it, called
Go, or AlphaGo, something like that.

1051
00:55:51,960 --> 00:55:53,440
Yeah, highly recommended.

1052
00:55:53,440 --> 00:55:53,970
Super--

1053
00:55:53,970 --> 00:55:59,750
KEVIN XU: Yes, so, furthermore, like
in terms of not just video games,

1054
00:55:59,750 --> 00:56:03,140
people are still developing things to--

1055
00:56:03,140 --> 00:56:09,240
just test what we can use this
for, in terms of not just these--

1056
00:56:09,240 --> 00:56:11,750
so, games have different
classifications,

1057
00:56:11,750 --> 00:56:15,750
and Carnegie Mellon actually
solved poker a few years ago.

1058
00:56:15,750 --> 00:56:20,370
Which is crazy, because poker, you
don't always have all the information.

1059
00:56:20,370 --> 00:56:22,460
So there's so much more
extrapolation that you

1060
00:56:22,460 --> 00:56:25,020
need to do from a given
subset of information.

1061
00:56:25,020 --> 00:56:30,050
And, so, it's actually very interesting
how far that machine learning

1062
00:56:30,050 --> 00:56:34,640
has come, from being able to take such a
small, comparatively small set of data,

1063
00:56:34,640 --> 00:56:40,540
and be able to generalize
it to big things.

1064
00:56:40,540 --> 00:56:43,270
SPEAKER: OK, and [? Yashvi ?]
asks, "Just as you mentioned,

1065
00:56:43,270 --> 00:56:46,030
we trained the computer
for game 1,000 times.

1066
00:56:46,030 --> 00:56:50,027
Was that the training part or the
testing part from train-test?"

1067
00:56:50,027 --> 00:56:51,110
KEVIN XU: Ah, yeah, great.

1068
00:56:51,110 --> 00:56:56,740
So, when we did reinforcement
learning, the training over 1,000 times

1069
00:56:56,740 --> 00:57:00,520
was simply to train the computer.

1070
00:57:00,520 --> 00:57:04,090
When you're reinforcement
learning, your testing part,

1071
00:57:04,090 --> 00:57:07,090
you don't really have
a test set, as opposed

1072
00:57:07,090 --> 00:57:11,020
to when you have a big data set of
images where you can divide 80% of them

1073
00:57:11,020 --> 00:57:15,070
to train it, and then verify that
your algorithm, or your regression

1074
00:57:15,070 --> 00:57:16,990
is correct with the remainder 20%.

1075
00:57:16,990 --> 00:57:21,400
In the case of the game, you just have--
like, once you've trained your AI,

1076
00:57:21,400 --> 00:57:22,960
you just have your AI play the game.

1077
00:57:22,960 --> 00:57:25,660
In which case, this
is your testing phase.

1078
00:57:25,660 --> 00:57:27,070
It's like, did the AI win?

1079
00:57:27,070 --> 00:57:31,280
Like in our case, it was like, how
many points did the AI get on average?

1080
00:57:31,280 --> 00:57:35,080
And, so, that is the testing
phase, which doesn't actually

1081
00:57:35,080 --> 00:57:38,750
use a data set but rather
just uses the game itself.

1082
00:57:38,750 --> 00:57:39,950
SPEAKER: OK.

1083
00:57:39,950 --> 00:57:43,940
"What is the--" This is from [? Madhi ?]
again, "What is the application--"

1084
00:57:43,940 --> 00:57:47,250
Oh, no I already asked
that one, sorry about that.

1085
00:57:47,250 --> 00:57:53,210
OK from HW, "Besides Google
TensorFlow, what other ML platforms

1086
00:57:53,210 --> 00:57:54,080
should we look into?

1087
00:57:54,080 --> 00:57:57,320
Anything from IBM or Microsoft?"

1088
00:57:57,320 --> 00:58:00,080
I think you answered that one, right?

1089
00:58:00,080 --> 00:58:00,880
KEVIN XU: It's--

1090
00:58:00,880 --> 00:58:02,105
OK, so with the--

1091
00:58:02,105 --> 00:58:07,100
the thing about libraries,
like, choosing the best library

1092
00:58:07,100 --> 00:58:08,550
can be very hard.

1093
00:58:08,550 --> 00:58:11,810
And this is like something that
requires knowledge, and like,

1094
00:58:11,810 --> 00:58:14,970
you to just jump into ML and see
what works and see what doesn't.

1095
00:58:14,970 --> 00:58:20,360
So, what we recommend, is that
you don't concentrate so much

1096
00:58:20,360 --> 00:58:23,930
on optimizing, right now, everything.

1097
00:58:23,930 --> 00:58:27,600
You want to get something that works
with the knowledge and the skills

1098
00:58:27,600 --> 00:58:31,400
that you have, first, before you
can improve and try and optimize

1099
00:58:31,400 --> 00:58:34,890
to get algorithms that are very good
for your data set and things like that.

1100
00:58:34,890 --> 00:58:36,443
So--

1101
00:58:36,443 --> 00:58:38,360
ZAD CHIN: I think the
other thing about trying

1102
00:58:38,360 --> 00:58:42,020
to find an ML library is to find
whether it is well documented

1103
00:58:42,020 --> 00:58:45,080
or whether actually a lot of
people have tried it as well.

1104
00:58:45,080 --> 00:58:47,780
I personally don't know any library.

1105
00:58:47,780 --> 00:58:51,410
Maybe I'm just shallow minded, but
I don't really know any library

1106
00:58:51,410 --> 00:58:52,970
from like IBM and Microsoft.

1107
00:58:52,970 --> 00:58:56,960
But I think they have a really strong
IBM and Microsoft Research team.

1108
00:58:56,960 --> 00:58:58,697
Which I really admire as well.

1109
00:58:58,697 --> 00:59:01,530
But in terms of TensorFlow, they
are really, really well documented.

1110
00:59:01,530 --> 00:59:05,005
There's a lot of examples out there, a
lot of Stack Overflow posts about that

1111
00:59:05,005 --> 00:59:06,380
which is actually very important.

1112
00:59:06,380 --> 00:59:08,517
But you actually face
a bottleneck, right?

1113
00:59:08,517 --> 00:59:11,600
Because you need someone to help you,
someone to understand your data set.

1114
00:59:11,600 --> 00:59:14,540
And everybody is using it, it
was really well documented.

1115
00:59:14,540 --> 00:59:17,870
So I would recommend TensorFlow
if you are beginning,

1116
00:59:17,870 --> 00:59:19,970
but like, if you want to
be more sophisticated,

1117
00:59:19,970 --> 00:59:22,380
you can start to build
your own neural network.

1118
00:59:22,380 --> 00:59:25,760
And I think that will be the most
sophisticated thing that you can do.

1119
00:59:25,760 --> 00:59:28,490
KEVIN XU: And of course, we
don't expect you guys, in CS50,

1120
00:59:28,490 --> 00:59:29,910
to build your own right now.

1121
00:59:29,910 --> 00:59:33,410
But, in the future,
just as consideration,

1122
00:59:33,410 --> 00:59:37,280
there's a lot of statistical theory
that goes into what machine learning is.

1123
00:59:37,280 --> 00:59:40,730
And, so, there are Harvard
classes, or MIT classes,

1124
00:59:40,730 --> 00:59:45,900
that you can take that are just
totally on ML and applying it.

1125
00:59:45,900 --> 00:59:48,480
And if you are interested
in talking about that,

1126
00:59:48,480 --> 00:59:52,135
feel free to reach out to one of us,
and we'd be happy to chat with you.

1127
00:59:52,135 --> 00:59:52,760
ZAD CHIN: Yeah.

1128
00:59:52,760 --> 00:59:56,960
And, so, before we actually end, please
take one minute to actually fill out

1129
00:59:56,960 --> 01:00:00,890
this feedback form, which is
tinyurl.com/ML50-feedback.

1130
01:00:00,890 --> 01:00:03,865
And we actually really thank
you so much for actually coming,

1131
01:00:03,865 --> 01:00:06,740
it's really late, or really early,
I don't know what's the time zone.

1132
01:00:06,740 --> 01:00:08,490
But thank you so much
for actually coming,

1133
01:00:08,490 --> 01:00:12,448
we actually enjoyed this great seminar,
and we are very excited to share this.

1134
01:00:12,448 --> 01:00:14,240
And we hope that you
learned something too.

1135
01:00:14,240 --> 01:00:14,740
So--

1136
01:00:14,740 --> 01:00:17,150


1137
01:00:17,150 --> 01:00:20,540
KEVIN XU: So, yeah, I think we'll
be ending the recording now,

1138
01:00:20,540 --> 01:00:26,200
but might stay a little bit after
to answer any further questions.

1139
01:00:26,200 --> 01:00:27,468