1
00:00:01,690 --> 00:00:10,350
Here we have validation and test data, it will set aside some samples for validation and testing during

2
00:00:10,430 --> 00:00:17,440
our training process as we are going to divide over samples, we have a total of 20 samples which will

3
00:00:17,440 --> 00:00:19,210
be divided into three sets.

4
00:00:19,390 --> 00:00:22,090
One set is for training the network.

5
00:00:22,330 --> 00:00:25,790
We need to give some problem and some solution.

6
00:00:25,840 --> 00:00:29,000
We are also showing the solution to our network.

7
00:00:29,260 --> 00:00:36,010
This just looked like when we are teaching in a classroom and we will give some example to students

8
00:00:36,160 --> 00:00:41,580
and we solve those examples and we will try to do it as much as we can here.

9
00:00:41,590 --> 00:00:42,940
It's 14 samples.

10
00:00:42,950 --> 00:00:46,750
We will do it for 70 percent of data and 14 samples.

11
00:00:46,750 --> 00:00:49,210
We can't really change the 70 percent.

12
00:00:49,660 --> 00:00:52,750
But here we have two more option validation and testing.

13
00:00:53,260 --> 00:00:58,900
Testing, as the name explain itself is testing like final exam.

14
00:00:59,200 --> 00:01:05,230
At the end of the semester, we will conduct a final examination for a student to check it out.

15
00:01:05,230 --> 00:01:11,650
If this training either if this course was useful, did the students learn something or not?

16
00:01:11,870 --> 00:01:15,790
We will put 15 percent of data for over testing.

17
00:01:15,940 --> 00:01:23,230
However, we can just adjust these numbers based on our data and training a strategy.

18
00:01:23,480 --> 00:01:28,840
I'm going to get to them later, but testing will happen at the end of the semester.

19
00:01:29,320 --> 00:01:35,310
And if 90 percent of the students fail, we cannot do anything.

20
00:01:35,980 --> 00:01:40,360
We did something wrong during the training, but it's already too late.

21
00:01:40,360 --> 00:01:47,390
And phrase that we use is training students or you see this next have been overtrained.

22
00:01:47,980 --> 00:01:54,520
What horrible validation here at the end of the semester, if we realize 90 percent of the students

23
00:01:54,520 --> 00:01:56,820
fail, it's already too late.

24
00:01:57,130 --> 00:02:06,040
What is strategy to understand and to find a problem earlier is conducting some quizzes during the semester,

25
00:02:06,220 --> 00:02:08,260
which we call them here validation.

26
00:02:08,260 --> 00:02:14,710
So validation are some testing and some quizzes during the training process.

27
00:02:14,950 --> 00:02:23,260
These are used to measure network generalization and to how training visualization a stop improving

28
00:02:23,470 --> 00:02:24,790
during the quizzes.

29
00:02:24,790 --> 00:02:32,470
If we realize that each quiz result is just worse than the previous one, we need to give up on those

30
00:02:32,470 --> 00:02:33,310
students.

31
00:02:33,400 --> 00:02:37,870
We need to change the training to agree to training a strategy.

32
00:02:38,230 --> 00:02:43,900
And there is no point for continue this training because we are getting the negative result.

33
00:02:44,140 --> 00:02:46,210
We shall stop the training.

34
00:02:46,780 --> 00:02:49,420
Let me just explain it with a better example.

35
00:02:54,500 --> 00:02:57,270
Let's take a look at these two networks.

36
00:02:58,200 --> 00:03:03,940
I'm going to choose color of blue for showing the result of training.

37
00:03:04,370 --> 00:03:05,930
This is our era.

38
00:03:06,230 --> 00:03:07,030
This is over.

39
00:03:07,040 --> 00:03:08,410
And just showing it with E!

40
00:03:08,780 --> 00:03:09,980
This is the error.

41
00:03:09,980 --> 00:03:14,300
And this is over time during the time first.

42
00:03:14,500 --> 00:03:15,500
This is the error.

43
00:03:15,500 --> 00:03:24,320
And then in this system, in this net for the error is going to almost zero, very close to zero.

44
00:03:24,720 --> 00:03:33,210
Our goal is to make the error zero, which means the system is a very good system and the error is zero.

45
00:03:33,770 --> 00:03:41,270
But in this system, in other hand, the error having to start from here and the same point, it's going

46
00:03:41,270 --> 00:03:44,620
towards zero, but not very good samples.

47
00:03:44,630 --> 00:03:46,580
It's going to be something like that.

48
00:03:47,240 --> 00:03:51,690
And then I'm going to show the validation data with the color of green.

49
00:03:52,160 --> 00:03:55,220
So this is my validation data.

50
00:03:55,580 --> 00:03:58,070
The validation will start from this error.

51
00:03:58,100 --> 00:04:01,350
This amount of error and you wouldn't get time is just improving.

52
00:04:01,370 --> 00:04:02,060
Very good.

53
00:04:02,390 --> 00:04:03,710
Oh, what happened here?

54
00:04:03,740 --> 00:04:05,900
It's just not going well.

55
00:04:05,930 --> 00:04:12,680
Now, the error here at this point is start to increasing.

56
00:04:13,340 --> 00:04:15,710
So it's not a very good team to training.

57
00:04:15,720 --> 00:04:16,720
Something went wrong.

58
00:04:16,730 --> 00:04:18,320
The training was good at the beginning.

59
00:04:18,320 --> 00:04:21,040
They were start to go towards zero.

60
00:04:21,050 --> 00:04:29,380
But now the error is not very good, meaning these students or this network is getting over trained.

61
00:04:30,530 --> 00:04:35,470
But on the other hand, in this network here, we have the validation data.

62
00:04:35,600 --> 00:04:40,760
It's just going good and the other one is over testing.

63
00:04:40,760 --> 00:04:48,620
You're going to just show it with a color of red to match the Matlab fine for the testing of this one.

64
00:04:48,620 --> 00:04:50,450
The testing is going actually good.

65
00:04:51,800 --> 00:04:59,360
We might get a good result for testing, but validation is not really good and we can't really trust

66
00:04:59,360 --> 00:04:59,500
it.

67
00:04:59,510 --> 00:05:03,610
And in this network, this is, let's say, our testing.

68
00:05:05,900 --> 00:05:10,190
So if you want to choose a product with this result, which one would you choose?

69
00:05:10,670 --> 00:05:12,860
The first one or the second one?

70
00:05:14,790 --> 00:05:16,320
That's what is happening here.

71
00:05:16,700 --> 00:05:25,200
Suppose we have two products, they are cooling systems, one of them has an error of minus plus 12

72
00:05:25,200 --> 00:05:26,250
percent error.

73
00:05:28,580 --> 00:05:39,250
This war is giving us an error around to person, which is very good, minus or plus two percent error,

74
00:05:39,260 --> 00:05:42,760
but sometimes it has eighty five percent error.

75
00:05:43,670 --> 00:05:52,130
So which product would you choose if you want to set your temperature to plus six?

76
00:05:53,180 --> 00:05:59,300
This network will give you something around six wheat, 12 percent of error.

77
00:05:59,720 --> 00:06:06,500
This system is giving you a better result in most of the time, but sometimes it might have 80 percent

78
00:06:06,950 --> 00:06:07,470
error.

79
00:06:07,820 --> 00:06:12,170
This part, of course, we cannot have this product.

80
00:06:12,170 --> 00:06:17,510
We cannot trust it because the behavior of the system is not predictable.

81
00:06:17,520 --> 00:06:19,860
We cannot predict the behavior of the system.

82
00:06:20,150 --> 00:06:22,760
That's why we can't really trust this network.

83
00:06:23,000 --> 00:06:29,480
And this network has the better result for us and we can trust this product more.

84
00:06:30,290 --> 00:06:35,300
So meaning these steps forward is just showing the better meniscal error.

85
00:06:35,300 --> 00:06:42,530
It has a better miniskirt error, but in a matter of demoralisation, it's not doing so well.

86
00:06:44,090 --> 00:06:44,640
The product.

87
00:06:44,660 --> 00:06:52,370
Number two, regarding the meter squared error, it's not as good as product one, but it's trustworthy.

88
00:06:52,370 --> 00:06:56,400
We can trust it because it's not going to change its behavior.

89
00:06:57,350 --> 00:07:03,440
This is a very important point that we need to keep in mind while we are training and we're all network.

90
00:07:03,620 --> 00:07:05,480
Now, let's back to Matlab.