1
00:00:00,450 --> 00:00:03,390
So now we take a look at regularisation methods.

2
00:00:03,750 --> 00:00:07,020
These are methods that help us prevent or reduce overfitting.

3
00:00:07,560 --> 00:00:08,610
So let's get started.

4
00:00:08,790 --> 00:00:14,940
So firstly, as I said, regularization is a technique that helps attempt to reduce overfitting and

5
00:00:14,940 --> 00:00:18,450
basically improve the generalization of your model.

6
00:00:18,960 --> 00:00:24,780
It allows us to complete control the model complexity so that we ensure the model is using the right

7
00:00:24,780 --> 00:00:26,280
features to classify the object.

8
00:00:26,640 --> 00:00:31,500
As you can see in this example, here we have we previously looked at we don't want a model that is

9
00:00:31,500 --> 00:00:34,260
learning that the green grass belongs to a dog.

10
00:00:34,650 --> 00:00:39,600
When all these images have dogs in them, we want them to learn what a teal knows is and motors.

11
00:00:39,990 --> 00:00:41,400
So there are many ways we can do that.

12
00:00:41,400 --> 00:00:45,830
So we'll take a look at these methods now basically in computer vision.

13
00:00:45,900 --> 00:00:52,230
These are the most common regularization methods that we use can use L1 L2 regularization.

14
00:00:52,710 --> 00:00:58,620
You can use data augmentation, which is something I actually use quite a bit in my deep learning models.

15
00:00:59,160 --> 00:01:01,950
You can use drop, which is something I actually always use.

16
00:01:02,130 --> 00:01:04,410
It's a very easy to implement, at least stop.

17
00:01:04,560 --> 00:01:07,050
It's essentially a regularization technique per say.

18
00:01:07,050 --> 00:01:07,980
It's just a good.

19
00:01:09,330 --> 00:01:14,490
It's a way of knowing when to stop creating a model so that we actually don't continue to over fit on

20
00:01:14,490 --> 00:01:16,260
data called early stopping.

21
00:01:16,770 --> 00:01:21,810
And then there's another technique that we can use in between layers called batch normalization.

22
00:01:22,470 --> 00:01:28,620
So it will stop there and we'll go into each of these techniques in detail or bit of detail.

23
00:01:29,010 --> 00:01:32,460
So we'll first start with L1 L2 regularization.

24
00:01:32,970 --> 00:01:38,190
When we've gone through all of these techniques, we'll go back to a code and we'll train.

25
00:01:38,190 --> 00:01:43,430
Some models firstly, will train them without any regularization and then we'll train them with what

26
00:01:43,770 --> 00:01:48,900
with to regularization a few of the techniques combined and we can compare their performance.

27
00:01:49,260 --> 00:01:55,500
So I'll see you in the next lesson where we take a look at L1 L2 regularization.