1
00:00:00,833 --> 00:00:03,066
Hello and welcome back to the course
on Machine Learning.

2
00:00:03,066 --> 00:00:05,966
Today we're talking about random forests.

3
00:00:05,966 --> 00:00:07,466
Let's have a look inside.

4
00:00:07,466 --> 00:00:10,000
So first thing we're going to
learn is a new concept.

5
00:00:10,000 --> 00:00:12,900
Ensemble learning.
What is ensemble learning.

6
00:00:12,900 --> 00:00:17,700
Ensemble learning is when you take
multiple machine learning algorithms

7
00:00:17,700 --> 00:00:21,233
and put them together to create one bigger
machine learning algorithm.

8
00:00:21,233 --> 00:00:24,800
So that machine learning algorithm,
the final one is actually using

9
00:00:24,800 --> 00:00:28,066
a leveraging many different other machine
learning algorithms.

10
00:00:28,066 --> 00:00:31,066
And they can be the same machine
learning algorithm as we will see today.

11
00:00:31,400 --> 00:00:32,633
We're going to be looking at the random

12
00:00:32,633 --> 00:00:36,366
forest method which combines
lots of decision tree methods.

13
00:00:36,366 --> 00:00:40,666
So instead of running a decision tree
method, which we talked about earlier,

14
00:00:40,666 --> 00:00:44,733
instead of running it once,
we're going to run it multiple times.

15
00:00:44,733 --> 00:00:46,033
And that will give us a random forest.

16
00:00:46,033 --> 00:00:49,166
So let's have a look at the step
by step process to understand

17
00:00:49,166 --> 00:00:50,700
how all of this happens.

18
00:00:50,700 --> 00:00:54,600
So step one is you pick at random
k data points from the training set.

19
00:00:55,100 --> 00:00:59,000
Then you build a decision tree
associated to those k data points.

20
00:00:59,333 --> 00:01:00,766
So instead of building a decision tree

21
00:01:00,766 --> 00:01:03,000
based on everything you have in
your data set,

22
00:01:03,000 --> 00:01:06,666
you build a decision tree just based on
some of the data points that you have.

23
00:01:06,666 --> 00:01:09,666
So kind of like a subset
of all of your data set.

24
00:01:09,666 --> 00:01:13,066
Next, you choose the number of trees
you want to build, and you repeat steps

25
00:01:13,066 --> 00:01:14,200
one and two.

26
00:01:14,200 --> 00:01:18,066
And then once you have all of those trees
and you have a new data point.

27
00:01:18,066 --> 00:01:20,100
So when you want to check where
and you thought about falls

28
00:01:20,100 --> 00:01:24,600
or how is classified for a new data point,
you make each one of your entry trees

29
00:01:24,600 --> 00:01:29,433
predict the category of which the data
point belongs to, and then you assign

30
00:01:29,433 --> 00:01:32,433
the new data point to the category
that wins the majority vote.

31
00:01:32,600 --> 00:01:34,766
So that's how a random forest works.

32
00:01:34,766 --> 00:01:38,233
So basically you start off with one tree
and then you build another tree,

33
00:01:38,233 --> 00:01:39,100
another tree, another tree.

34
00:01:39,100 --> 00:01:40,500
And each one of those trees

35
00:01:40,500 --> 00:01:43,900
is being built on a randomly selected
subset from your data.

36
00:01:44,333 --> 00:01:47,866
And even though each one of those trees
might not be ideal

37
00:01:48,033 --> 00:01:50,800
overall, on average
they can perform very well.

38
00:01:50,800 --> 00:01:53,533
And that's a major advantage
of this algorithm.

39
00:01:53,533 --> 00:01:57,100
It's kind of leveraging
the power of the crowd, so to speak,

40
00:01:57,100 --> 00:01:58,800
instead of just relying on one tree.

41
00:01:58,800 --> 00:02:01,766
It's checking
what all the trees are going to say,

42
00:02:01,766 --> 00:02:05,600
and then just taking the majority vote
and deciding the classification

43
00:02:05,600 --> 00:02:10,800
based on that and that power of numbers
can help get rid of certain errors and

44
00:02:10,800 --> 00:02:14,766
certain uncertainties in your algorithm
and make it more precise.

45
00:02:15,333 --> 00:02:18,900
And in fact, it's such a good solution
that when Microsoft were developing

46
00:02:18,900 --> 00:02:23,400
Kinect, you know, this device that allows
you to play games on your television,

47
00:02:23,733 --> 00:02:26,733
that little device over there,
it attaches to Xbox,

48
00:02:26,900 --> 00:02:29,100
and then you can play games
without any controller.

49
00:02:29,100 --> 00:02:32,700
So here
that device is using an infrared grid

50
00:02:32,900 --> 00:02:37,233
to understand where the hands, arms, heads

51
00:02:37,233 --> 00:02:40,533
and other parts of the body of these
people are located and how they're moving.

52
00:02:41,033 --> 00:02:44,700
And it's using machine
learning to understand how

53
00:02:44,700 --> 00:02:48,333
the body parts are moving and where
exactly they're located in space.

54
00:02:48,666 --> 00:02:53,500
So when Microsoft was developing Kinect,
they decided to go with the Random Forest

55
00:02:53,500 --> 00:02:57,166
algorithm over all of the other machine
learning algorithms

56
00:02:57,166 --> 00:03:01,600
that were available to them,
and use the random forest to develop

57
00:03:01,833 --> 00:03:04,800
this sophisticated piece of hardware.

58
00:03:04,800 --> 00:03:05,933
Software.

59
00:03:05,933 --> 00:03:08,666
And actually,
they have a interesting article about it.

60
00:03:08,666 --> 00:03:09,833
So I'm just going to show it to you.

61
00:03:09,833 --> 00:03:12,766
Now with you,
you can find on the internet.

62
00:03:12,766 --> 00:03:14,833
So it's at microsoft.com.

63
00:03:14,833 --> 00:03:16,500
This is from there.

64
00:03:16,500 --> 00:03:18,900
And you can definitely find it there.

65
00:03:18,900 --> 00:03:21,800
It's called real time
human pose recognition parts

66
00:03:21,800 --> 00:03:23,766
from single depth images.

67
00:03:23,766 --> 00:03:25,666
And here it explains exactly.

68
00:03:25,666 --> 00:03:29,633
So you can actually see
the random forest in action.

69
00:03:29,633 --> 00:03:30,800
So you can see that

70
00:03:30,800 --> 00:03:34,400
this is similar to what we were talking
about before in the classification trees.

71
00:03:34,400 --> 00:03:36,300
But here it's actually
using random forests

72
00:03:36,300 --> 00:03:39,166
to understand where body parts are.

73
00:03:39,166 --> 00:03:42,133
And then based on that the device

74
00:03:42,133 --> 00:03:45,266
finds, 
what it needs to do in that computer game.

75
00:03:45,700 --> 00:03:47,100
So that's how it works.

76
00:03:47,100 --> 00:03:48,033
And so here,

77
00:03:48,033 --> 00:03:52,266
if I just search for the word forest,
you'll see that, decision forest.

78
00:03:52,266 --> 00:03:53,966
Decision forest. The decision forest.

79
00:03:53,966 --> 00:03:58,000
And they actually explained that,
they were able to achieve faster speeds,

80
00:03:58,366 --> 00:04:01,966
faster processing with the decision forest
and therefore that, you know,

81
00:04:01,966 --> 00:04:05,166
reduced the cost of the hardware
that they required for this tool.

82
00:04:05,700 --> 00:04:07,033
interesting article.

83
00:04:07,033 --> 00:04:07,500
check it out.

84
00:04:07,500 --> 00:04:10,966
If you want to learn a bit more
about a real life practical application

85
00:04:10,966 --> 00:04:14,000
of a decision forest or a random forest.

86
00:04:14,266 --> 00:04:15,800
And that's it for today's tutorial.

87
00:04:15,800 --> 00:04:20,333
I hope you enjoyed learning a bit
about an ensemble type of machine

88
00:04:20,333 --> 00:04:21,133
learning algorithm.

89
00:04:21,133 --> 00:04:24,366
Definitely the practical side
is going to be quite interesting as well,

90
00:04:24,666 --> 00:04:26,266
and I look forward
to seeing you next time.

91
00:04:26,266 --> 00:04:27,966
Until then, enjoy machine learning.