﻿1
00:00:00,333 --> 00:00:02,666
Hello and welcome back to the course
on Machine Learning.

2
00:00:02,666 --> 00:00:07,233
In today's tutorial, we will find out
how we can take our Non-Linearly separable

3
00:00:07,233 --> 00:00:12,033
data set, map it to a higher dimension
and get a linearly separable data set.

4
00:00:12,200 --> 00:00:17,000
Invoke the support vector machine
algorithm, build a decision boundary

5
00:00:17,000 --> 00:00:21,000
for a data set, and then project all of
that back into our original dimensions.

6
00:00:21,266 --> 00:00:24,266
So quite a lot to cover.
Let's get started.

7
00:00:24,600 --> 00:00:26,933
First off we're going to
look at a simplified example.

8
00:00:26,933 --> 00:00:29,600
We're going to look
at a one dimensional data set.

9
00:00:29,600 --> 00:00:33,833
So normally we visualize everything
in the PowerPoint with two dimensions.

10
00:00:34,066 --> 00:00:36,633
to make it, you know, look pretty.

11
00:00:36,633 --> 00:00:40,466
And so that we can, kind of understand
how it would work in multiple dimensions.

12
00:00:40,633 --> 00:00:44,333
But right now, that will be a bit
too complex for us to start with.

13
00:00:44,366 --> 00:00:47,300
So we're going to start
with a single dimension.

14
00:00:47,300 --> 00:00:49,333
So here we've got the X1 dimension.

15
00:00:49,333 --> 00:00:52,533
We've got some points here.

16
00:00:52,533 --> 00:00:54,100
So we've got nine data points.

17
00:00:54,100 --> 00:00:57,500
And as we can see
they are non linearly separable.

18
00:00:57,500 --> 00:01:01,433
So in a in this dimensional 
in a single dimension

19
00:01:02,066 --> 00:01:06,333
dimensional space, a linear separator
would not be a line would be a dot.

20
00:01:06,333 --> 00:01:06,633
Right.

21
00:01:06,633 --> 00:01:10,233
So in a two dimensional space,
a linear, a linear

22
00:01:10,700 --> 00:01:15,700
separator is a line, in a three
dimensional space it's a hyperplane.

23
00:01:15,700 --> 00:01:18,133
But in a one dimensional space
it's a single dot.

24
00:01:18,133 --> 00:01:21,266
So can we separate a separate
the green from the red with a single dot?

25
00:01:21,266 --> 00:01:21,700
Here.

26
00:01:21,700 --> 00:01:25,833
No we cannot we if we put it here
then these will be separated from that.

27
00:01:25,833 --> 00:01:27,633
If we put it here
will be separate from that.

28
00:01:27,633 --> 00:01:30,633
So this is a non-linearly separable
data set.

29
00:01:31,300 --> 00:01:34,033
Now how can we apply

30
00:01:34,033 --> 00:01:38,400
the method of increasing
the dimensionality of this space

31
00:01:38,400 --> 00:01:42,066
to make it a linearly separable data
set in a higher dimension?

32
00:01:42,066 --> 00:01:43,566
That's what we're going to look at.

33
00:01:43,566 --> 00:01:46,033
And it might seem impossible. Right.

34
00:01:46,033 --> 00:01:49,766
So the first time I learned about this,
for me it was like, wow, how can you take,

35
00:01:50,300 --> 00:01:54,600
a non-linearly separable data set
and somehow magically increase

36
00:01:54,600 --> 00:01:58,566
the dimensionality
and you get a, linearly separable data

37
00:01:58,566 --> 00:02:02,833
set that, you know, sounded like absurd,
but it is actually possible.

38
00:02:02,833 --> 00:02:05,300
And that's
what we're going to see right now.

39
00:02:05,300 --> 00:02:07,900
So we're going to create this
mapping function on the fly.

40
00:02:07,900 --> 00:02:13,033
So, let's say that
this point over here is around five.

41
00:02:13,033 --> 00:02:14,833
So, this is zero over here.

42
00:02:14,833 --> 00:02:17,600
And then somewhere here we've got five.
It doesn't really matter.

43
00:02:17,600 --> 00:02:18,333
It can be any number.

44
00:02:18,333 --> 00:02:22,433
But just for argument's sake, let's say
that this point over here is five, right?

45
00:02:22,433 --> 00:02:23,233
And then it keeps going.

46
00:02:23,233 --> 00:02:26,066
So our first step to build
the mapping function,

47
00:02:26,066 --> 00:02:28,533
there can be multiple mapping functions
that you can build.

48
00:02:28,533 --> 00:02:31,533
I'm just going to show you
one that came to my mind.

49
00:02:31,600 --> 00:02:35,400
so the first step will be to go
f equals x minus five.

50
00:02:35,500 --> 00:02:38,966
So to subtract five from our data set.

51
00:02:38,966 --> 00:02:43,066
And that is going to 
what is that going to do.

52
00:02:43,066 --> 00:02:46,966
That is going to 
move everything to the left.

53
00:02:46,966 --> 00:02:52,500
So basically, now,
this is what, the result looks like.

54
00:02:52,500 --> 00:02:56,533
So if you take five, you subtract
five from x, you get, you know,

55
00:02:56,533 --> 00:03:00,000
like these ones will go into negative,
these ones will stay and positive.

56
00:03:00,900 --> 00:03:04,266
And then the next step
would be to, square all of that.

57
00:03:04,266 --> 00:03:07,766
So f is now equals to x minus five square.

58
00:03:07,766 --> 00:03:09,066
So how will that all look like.

59
00:03:09,066 --> 00:03:13,066
Well basically
you'll have this squared function

60
00:03:13,066 --> 00:03:17,033
going through your, chart.

61
00:03:17,033 --> 00:03:20,033
And then all of these will be projected
onto the function.

62
00:03:20,266 --> 00:03:21,166
there we go.

63
00:03:21,166 --> 00:03:24,266
So that's what it looks like
f equals x minus five squared.

64
00:03:24,633 --> 00:03:28,133
And now what we want to do
is we just want to see

65
00:03:28,133 --> 00:03:29,966
that it is indeed linearly separable.

66
00:03:29,966 --> 00:03:30,800
So there we go.

67
00:03:30,800 --> 00:03:32,933
There's our linear separator.

68
00:03:32,933 --> 00:03:36,200
So in a two dimensional space
as we remember a linear separator

69
00:03:36,200 --> 00:03:37,766
is a straight line.

70
00:03:37,766 --> 00:03:40,766
And as you can see this data set

71
00:03:40,966 --> 00:03:43,733
became linearly separable
in this dimension.

72
00:03:43,733 --> 00:03:46,566
I know it's it's surprising
and even a bit shocking.

73
00:03:46,566 --> 00:03:48,200
But indeed it is is the case.

74
00:03:48,200 --> 00:03:52,400
So you can see that we were able to take
this line and separate all of them,

75
00:03:52,700 --> 00:03:57,900
red elements or other data set
from the green elements, and that's it.

76
00:03:58,200 --> 00:04:01,500
And then what we would do next from here
is we would project everything back

77
00:04:01,500 --> 00:04:05,433
onto our original space,
and we would know how

78
00:04:05,433 --> 00:04:09,666
to functionally separate,
the green from the red.

79
00:04:10,400 --> 00:04:15,000
And that is, what happens when you map
something to a higher dimension.

80
00:04:15,300 --> 00:04:20,500
So now knowing this example
and seeing that it works in, reality,

81
00:04:20,966 --> 00:04:24,566
we can proceed to a higher dimension,
you know, to start

82
00:04:24,566 --> 00:04:26,666
with a two dimensional space.
So let's have a look.

83
00:04:26,666 --> 00:04:28,233
So there's our two dimensional space.

84
00:04:28,233 --> 00:04:30,933
And basically
you would apply the same principle.

85
00:04:30,933 --> 00:04:35,666
So here you can apply you kind of invoke
the support vector machine algorithm

86
00:04:35,666 --> 00:04:38,933
because it is not a non, linearly

87
00:04:39,333 --> 00:04:42,766
separable data set in this space.

88
00:04:42,766 --> 00:04:45,833
But then you would apply
some sort of mapping function

89
00:04:45,833 --> 00:04:51,133
like right now we won't go into detail
what exactly mapping function it would be.

90
00:04:51,300 --> 00:04:54,133
And again there could be multiple
different options and so on.

91
00:04:54,133 --> 00:04:56,933
But basically
based on the previous example

92
00:04:56,933 --> 00:05:00,700
we now know that it's possible,
like we've seen, empirical evidence

93
00:05:00,700 --> 00:05:01,800
that it is possible

94
00:05:01,800 --> 00:05:03,433
to do the same thing applies

95
00:05:03,433 --> 00:05:05,866
to two dimensional space
moving into a three dimensional space,

96
00:05:05,866 --> 00:05:08,733
you would map it into a three
dimensional space, and then somehow

97
00:05:08,733 --> 00:05:13,666
it would become a linearly separable data
set in this space.

98
00:05:13,666 --> 00:05:16,200
And here we've got the new dimension,
which is Z.

99
00:05:16,200 --> 00:05:18,433
And in a three dimensional space,

100
00:05:18,433 --> 00:05:22,166
the linear separator is no longer
a line, it's a hyperplane.

101
00:05:22,400 --> 00:05:25,433
And so this hyperplane separates the two,

102
00:05:26,000 --> 00:05:29,000
parts of our data set in the way we want.

103
00:05:29,200 --> 00:05:33,400
So the support vector machine algorithm
has helped us build this hyperplane.

104
00:05:33,733 --> 00:05:36,400
And then basically
so we've got this result that

105
00:05:36,400 --> 00:05:39,700
we just projected back
into our two dimensional space.

106
00:05:39,700 --> 00:05:43,800
And we've got this, circle
that encompasses our,

107
00:05:44,666 --> 00:05:47,966
classes or separates our classes.

108
00:05:48,266 --> 00:05:49,433
And there we go.

109
00:05:49,433 --> 00:05:51,300
We've got the non linear separator.

110
00:05:51,300 --> 00:05:57,100
So as you can see
we can still even though we I've got a

111
00:05:57,433 --> 00:06:00,433
a bit of a more complex problem
where we cannot,

112
00:06:00,466 --> 00:06:04,200
directly apply the support
vector machine algorithm as we used to.

113
00:06:04,800 --> 00:06:09,400
we can still go into
a, higher, dimension

114
00:06:09,400 --> 00:06:13,633
and then apply the support
vector machine algorithm and, like,

115
00:06:13,633 --> 00:06:16,866
we won't go into detail
if it's possible all the time.

116
00:06:16,866 --> 00:06:18,533
And if this case is one,
it's not possible.

117
00:06:18,533 --> 00:06:19,633
And so what do you do there?

118
00:06:19,633 --> 00:06:23,666
But the point is that,
there is a solution that you can

119
00:06:23,866 --> 00:06:27,800
explore the, higher dimensions
so that this is not a dead end.

120
00:06:27,800 --> 00:06:30,266
You can just do that.
But there's a problem.

121
00:06:30,266 --> 00:06:33,500
The problem with this algorithm,
and the problem is that mapping

122
00:06:33,500 --> 00:06:36,600
to a high dimensional space
can be highly compute intensive.

123
00:06:36,600 --> 00:06:41,233
So it might require a lot of computation,
a lot of, processing power.

124
00:06:41,233 --> 00:06:46,400
And, you know, the larger your data set,
the more, of a problem this can cause.

125
00:06:47,366 --> 00:06:49,200
And therefore,

126
00:06:49,200 --> 00:06:52,333
this approach isn't the best
because you can imagine, like,

127
00:06:52,566 --> 00:06:56,966
you have a data set and then, 
mapping it to a higher dimension,

128
00:06:57,000 --> 00:07:00,166
performing all the calculations there,
and then coming back to your lower

129
00:07:00,166 --> 00:07:03,200
dimension that can,

130
00:07:03,200 --> 00:07:06,600
even for a computer,
not just like in our minds as humans,

131
00:07:06,800 --> 00:07:09,733
but just like for a computer that can,

132
00:07:09,733 --> 00:07:13,266
cause a lot of, delays.

133
00:07:13,266 --> 00:07:16,566
It can cause a lot of, like, processing

134
00:07:16,566 --> 00:07:19,666
backlog and issues in that sense.

135
00:07:19,666 --> 00:07:21,533
And, we don't want that to happen.

136
00:07:21,533 --> 00:07:22,966
And therefore, we're going to explore

137
00:07:22,966 --> 00:07:26,533
something else
we can explore, a a different approach,

138
00:07:26,766 --> 00:07:30,033
which is called in mathematics,
the kernel trick.

139
00:07:30,033 --> 00:07:34,500
And that's, approach
is going to help us perform

140
00:07:34,500 --> 00:07:38,700
very similar, gets very similar results,

141
00:07:38,700 --> 00:07:42,600
but without having to go
to a higher dimensional space.

142
00:07:42,600 --> 00:07:45,200
And we're going to talk about that
in the next

143
00:07:45,200 --> 00:07:47,933
tutorial, is going to be exciting,
and I can't wait to see you there.

144
00:07:47,933 --> 00:07:49,800
Until then, happy analyzing.
