1
00:00:00,450 --> 00:00:00,870
Hello.

2
00:00:00,900 --> 00:00:01,860
Welcome back.

3
00:00:01,860 --> 00:00:08,940
The cause of poor performance in machine learning can be attributed to either over 15 or under 15.

4
00:00:08,970 --> 00:00:15,920
What we want is a model that fits appropriately over fit in happens when a model learns the detail and

5
00:00:15,920 --> 00:00:23,160
the noise in the training data to the extent that it negatively impacts the performance of the model

6
00:00:23,220 --> 00:00:24,710
on new data.

7
00:00:24,720 --> 00:00:31,590
This means that the noise or random fluctuations in the training data is picked up and learned as a

8
00:00:31,590 --> 00:00:34,040
concept by the model.

9
00:00:34,050 --> 00:00:41,460
The problem is that these concepts do not apply to new data and negatively impact the modules ability

10
00:00:41,460 --> 00:00:49,110
to generalize and generalization refers to how well the concept learned by a machine learning model

11
00:00:49,320 --> 00:00:58,620
can apply to specific examples not seen by the module when it was learned the goal of a good machine

12
00:00:58,620 --> 00:01:06,120
learning module is to generalize well from the training data to a new data from the problem domain.

13
00:01:06,120 --> 00:01:12,090
This allows us to make predictions in the future on did on the data.

14
00:01:12,090 --> 00:01:14,560
The model has never seen before.

15
00:01:14,640 --> 00:01:21,590
In summary of our Fit and refers to a model that models the training data too well and under fits in

16
00:01:21,630 --> 00:01:28,020
on the other hand refers to a model that can need our model the training data no generalized to new

17
00:01:28,020 --> 00:01:28,530
data.

18
00:01:29,100 --> 00:01:37,830
Ideally we want to select a model at the sweet spot between under 15 and over 15.

19
00:01:37,830 --> 00:01:43,550
This is their goal but it's a bit difficult in practice to achieve without a bit of trial and error.

20
00:01:43,560 --> 00:01:50,130
An example of fitting is when your training set error is 1 percent and your test error is somewhere

21
00:01:50,130 --> 00:01:51,510
around 15 percent.

22
00:01:51,650 --> 00:01:56,970
Unlike sample 4 under fitness when you are training set error is 15 percent and your test set error

23
00:01:57,000 --> 00:02:04,860
is very large as well like 16 percent and humans are able to make the same prediction with a 0 percent

24
00:02:04,920 --> 00:02:05,540
error.

25
00:02:05,580 --> 00:02:13,380
1 in humans perform 15 percent better on the training set and 16 percent better on the test set.

26
00:02:13,530 --> 00:02:21,810
Another name for under 15 is high bias and another name for over 15 is high variance when under 15 occurs

27
00:02:21,900 --> 00:02:24,540
there are three things we can generally do.

28
00:02:24,540 --> 00:02:27,660
We can make our neural network bigger.

29
00:02:27,660 --> 00:02:29,610
We can train for a longer time.

30
00:02:29,850 --> 00:02:36,840
Oh we can also try a different neuro network architecture when over 15 years we can either get more

31
00:02:36,840 --> 00:02:45,660
data apply a method known as regularization or try different neural network architecture so that's all

32
00:02:45,660 --> 00:02:48,650
there is for this lesson and I shall see you in the next lesson.

33
00:02:48,660 --> 00:02:49,320
Have a nice day.