1
00:00:01,650 --> 00:00:08,650
By now you would have noticed that every architecture has in general two parts.

2
00:00:09,300 --> 00:00:14,420
The first part of the architecture is a conventional base.

3
00:00:14,610 --> 00:00:20,410
This includes conventional layers and pulling layers.

4
00:00:20,670 --> 00:00:24,420
It could be one conversational and one willingly.

5
00:00:24,450 --> 00:00:25,830
It could be tens of them.

6
00:00:25,830 --> 00:00:28,140
It could be hundreds of them.

7
00:00:28,140 --> 00:00:34,380
However if you look at the architectures most of them in the first part had convolution.

8
00:00:34,380 --> 00:00:41,990
Basically the output of that conventional base goes into a fully connected neural network.

9
00:00:43,500 --> 00:00:47,110
So the job of convolution base is very generic.

10
00:00:47,430 --> 00:00:53,570
It is to find out and highlight certain features from the input images.

11
00:00:53,850 --> 00:01:02,910
For example if you're inputting cats and dogs data the job of conventional base would be to highlight

12
00:01:03,060 --> 00:01:15,270
eyes ears whiskers claws etc. So all of these individual features of the image are highlighted by the

13
00:01:15,270 --> 00:01:17,520
Congressional base.

14
00:01:17,880 --> 00:01:25,530
The job of fully connected neural network is to use these identified features to classify the image

15
00:01:25,620 --> 00:01:30,400
whether it is a dog or whether it is a cat.

16
00:01:30,480 --> 00:01:40,080
So if you have a neural network which is already trained in identifying certain features and then classifying

17
00:01:40,080 --> 00:01:48,990
those images and now you have a new problem in which you are also trying to find the same features maybe

18
00:01:48,990 --> 00:01:51,440
you are trying to have a different classification.

19
00:01:51,570 --> 00:01:58,800
But if the import images have the same features in that case you can use these same convolution base

20
00:01:59,430 --> 00:02:03,030
of pre-teen models.

21
00:02:03,030 --> 00:02:13,170
For example in 2014 the AI unless we asked each challenge had one million images of different animals

22
00:02:14,280 --> 00:02:19,220
and there were 1000 different animals to which these images belong.

23
00:02:21,700 --> 00:02:30,940
The convolutions base of the winning networks what identifying features of different animals and declassified

24
00:02:30,990 --> 00:02:36,990
in the end were only classifying those features into which animal it is.

25
00:02:37,260 --> 00:02:43,010
And what is the breed of that animal.

26
00:02:43,020 --> 00:02:50,040
So now if we are only running a model to classify cats and dogs.

27
00:02:50,040 --> 00:02:54,730
This is the similar kind of input images that that particular talent had.

28
00:02:55,620 --> 00:03:03,350
So a model that is trained on 2014 data that can be used in our problem also.

29
00:03:06,130 --> 00:03:11,910
So this is the concept of crossover learning or feature extraction.

30
00:03:11,910 --> 00:03:19,560
We are going to use some part of our pre train model mostly the conventional base because conversational

31
00:03:19,560 --> 00:03:21,550
base is more genetic.

32
00:03:21,690 --> 00:03:30,120
It is only finding features and we will put a new classifier in front of the Congressional base that

33
00:03:30,140 --> 00:03:40,050
classifier will be trained by our system and that will be trained to classify and identify are images

34
00:03:40,380 --> 00:03:42,420
into the classes that we have.

35
00:03:44,700 --> 00:03:47,310
So this kind additional base will remain the same.

36
00:03:47,310 --> 00:03:53,130
We will have a new classified on top of it to classify our images.

37
00:03:53,130 --> 00:04:00,930
The advantages of doing this is that it saves a lot of time because we do not have to train this part

38
00:04:00,960 --> 00:04:03,240
of the network.

39
00:04:03,240 --> 00:04:06,720
Another good thing is these are proven models.

40
00:04:06,780 --> 00:04:10,490
They are one of the best in finding the features.

41
00:04:10,590 --> 00:04:18,330
So when we take their conversational base we can be assured that the features extracted from the images

42
00:04:18,720 --> 00:04:22,180
would be the best also.

43
00:04:22,420 --> 00:04:25,950
These models are trained on huge data set.

44
00:04:26,140 --> 00:04:30,720
They had input data of millions of images.

45
00:04:31,000 --> 00:04:38,170
So even if you have a small dataset from which featured extraction could have been difficult.

46
00:04:38,170 --> 00:04:47,980
These models are already trained to extract features on large amount of data and the best thing is that

47
00:04:47,980 --> 00:04:49,300
very easy to use.

48
00:04:49,380 --> 00:04:52,230
They are part of the Get Us library.

49
00:04:52,240 --> 00:04:59,350
It only takes a few lines of code to download all the weight of all the neurons in the econ relational

50
00:04:59,350 --> 00:05:01,720
base and those can be used straightaway.

51
00:05:02,890 --> 00:05:11,590
So in that project we will see how using pre-teen models we can achieve a higher level of accuracy even

52
00:05:11,590 --> 00:05:14,740
if we have small amount of data to not model.