1
00:00:00,390 --> 00:00:07,740
Here we have different plots, we have performance, we have training stage, we have our histogram

2
00:00:08,100 --> 00:00:10,330
and we have regression and feet.

3
00:00:11,070 --> 00:00:13,140
Let's take a look at this performance.

4
00:00:14,130 --> 00:00:15,990
Just click, OK?

5
00:00:17,460 --> 00:00:23,130
Here we can see the performance of our neural network at the beginning.

6
00:00:23,130 --> 00:00:26,240
The validation test result is just good.

7
00:00:26,520 --> 00:00:28,020
It's a good result.

8
00:00:28,020 --> 00:00:30,390
It's going toward zero.

9
00:00:30,420 --> 00:00:34,580
This is a verminous score forever and our goal is actually to go for zero.

10
00:00:34,830 --> 00:00:37,260
The training result actually is good here.

11
00:00:37,260 --> 00:00:39,540
It's just going for zero.

12
00:00:39,600 --> 00:00:44,630
But the problem was validation we had failing for validation.

13
00:00:44,630 --> 00:00:49,590
The validation just failed here once and then it just correct itself again.

14
00:00:49,800 --> 00:00:51,450
Improving, improving.

15
00:00:51,480 --> 00:00:52,650
And here it is.

16
00:00:52,650 --> 00:00:54,480
We have another fail.

17
00:00:54,660 --> 00:00:57,450
It will count the number of failures.

18
00:00:57,780 --> 00:01:02,400
If our system fails six times, then we will disrupt the training.

19
00:01:02,610 --> 00:01:10,620
But the former system for validation failed once here and then improve itself again will continue the

20
00:01:10,620 --> 00:01:13,920
validation and let us start six times.

21
00:01:13,920 --> 00:01:17,430
One, two, three, four, five and six.

22
00:01:17,700 --> 00:01:20,700
Then he just stopped the training process.

23
00:01:22,720 --> 00:01:28,160
Let's take a look at training a state and see what we have here.

24
00:01:28,780 --> 00:01:35,220
We have the gradient, we can see the mill, and this one is a validation check.

25
00:01:35,590 --> 00:01:42,160
So for the validation check here, we had some failure during the process, but it's just correct itself

26
00:01:42,160 --> 00:01:44,560
and try to improve again.

27
00:01:45,730 --> 00:01:48,380
We had a total of 17 EPOP.

28
00:01:48,790 --> 00:01:58,140
This means our system trained itself 17 times each time, change eight weights and then train again.

29
00:01:58,450 --> 00:02:01,850
And here are the results for failure.

30
00:02:01,870 --> 00:02:07,720
We had one failure here two times three, four, five and six.

31
00:02:08,170 --> 00:02:15,550
And I remember we just mentioned that if it just reached six times failure continuously, then I stopped

32
00:02:15,550 --> 00:02:16,840
the training process.

33
00:02:18,310 --> 00:02:20,230
Let's take a look at Şeref.

34
00:02:20,230 --> 00:02:27,400
Histogram error of histogram should be around zero, but we can see some of the testing are not good.

35
00:02:27,400 --> 00:02:30,330
They're not around zero.

36
00:02:30,340 --> 00:02:32,260
So it's not a good error.

37
00:02:32,260 --> 00:02:40,150
Histogram, if you have more data, if you have more samples, then you can have a better result and

38
00:02:40,150 --> 00:02:43,330
you can just understand it better.

39
00:02:45,780 --> 00:02:47,550
Let's check the regression.

40
00:02:48,330 --> 00:02:52,900
OK, here it is, we have a good training for regression.

41
00:02:53,730 --> 00:03:01,220
The goal here is why should we be equal to teach cheese over Target and why is our output?

42
00:03:01,650 --> 00:03:08,760
We can see the output here is almost equal one times the target plus this amount.

43
00:03:09,210 --> 00:03:12,850
And this is upper regression coefficient.

44
00:03:13,680 --> 00:03:15,620
These are our data.

45
00:03:15,660 --> 00:03:22,380
We can see the data here and it just try to feed to the data.

46
00:03:22,620 --> 00:03:24,140
And it was a good training.

47
00:03:24,270 --> 00:03:29,150
We also saw it on our performance that the training gets.

48
00:03:29,220 --> 00:03:30,120
It's doing good.

49
00:03:30,300 --> 00:03:33,450
But the problem was with validation and with testing.

50
00:03:33,690 --> 00:03:36,030
Testing actually is not good at all.

51
00:03:36,240 --> 00:03:42,420
We also saw it in the error histogram that testing is not around zero.

52
00:03:42,420 --> 00:03:45,240
So we can see here it's not a good fit.

53
00:03:45,450 --> 00:03:51,620
It should be toward this line and it should be equal to one for the validation.

54
00:03:51,630 --> 00:03:52,590
Here we have it.

55
00:03:52,590 --> 00:03:59,760
And this is showing the overall performance of our system, which is not bad, but not very good.

56
00:03:59,770 --> 00:04:02,880
Also here we have one data here, one data here.

57
00:04:03,210 --> 00:04:07,740
And for these two data, we don't have a very good fit.

58
00:04:07,920 --> 00:04:11,930
Maybe if we increase the number of samples, then we can have a better result.

59
00:04:13,260 --> 00:04:15,420
And finally, we have fit.

60
00:04:16,680 --> 00:04:19,640
This is a function feed or the output element.

61
00:04:19,950 --> 00:04:22,700
We know that our output is a sign of function.

62
00:04:23,010 --> 00:04:26,970
So these must be look like a sinus function.

63
00:04:26,970 --> 00:04:29,850
But however, it's not very good.

64
00:04:29,850 --> 00:04:36,930
It's almost like a sinus function, but not very good fit, meaning the neural network that we train

65
00:04:36,990 --> 00:04:43,860
was able to come up with this output for a sinus function, which is not very good.

66
00:04:43,860 --> 00:04:46,140
And also it's not a disaster.

67
00:04:46,140 --> 00:04:49,980
It's like sinus function, but not very good.

68
00:04:49,980 --> 00:04:52,110
Fit this part.

69
00:04:52,440 --> 00:05:00,030
The orange line is showing the error and we can see the error is too much here for these two samples.

70
00:05:00,030 --> 00:05:01,500
The error is a lot.

71
00:05:02,100 --> 00:05:04,590
Blue Dot is showing the training target.

72
00:05:05,070 --> 00:05:11,040
The plus is showing the training output and going for validation rate for testing.

73
00:05:11,280 --> 00:05:14,190
In this one is overfitting line.

74
00:05:14,670 --> 00:05:19,350
Here are our errors for the first two samples.

75
00:05:19,650 --> 00:05:20,970
It's actually a lot.