1
00:00:00,990 --> 00:00:08,220
Now, in this section, we are going to discuss the architecture of some popular CNN models, which

2
00:00:08,220 --> 00:00:11,110
one image classification competitions in the past.

3
00:00:12,990 --> 00:00:14,610
There are two reasons for doing this.

4
00:00:15,810 --> 00:00:23,160
First is we want to understand these architectures because this will build our intuition as to what

5
00:00:23,430 --> 00:00:25,490
is a good CNN model architecture.

6
00:00:27,390 --> 00:00:34,150
And the second reason is that these green models can be used by us in our software.

7
00:00:34,410 --> 00:00:43,410
Without retraining, these models with the architecture and the trained wait can be downloaded as part

8
00:00:43,410 --> 00:00:44,790
of the Kiraz Library only.

9
00:00:46,380 --> 00:00:53,880
But how will these train models, which are trained on other set, how can these be used for our classification

10
00:00:53,880 --> 00:00:54,240
problem?

11
00:00:55,800 --> 00:00:58,860
We will see the answer to this question in the coming lectures.

12
00:00:59,910 --> 00:01:08,070
Before we discuss these architectures, I briefly tell you about the image net competition which has

13
00:01:08,070 --> 00:01:10,440
given us these popular architectures.

14
00:01:11,580 --> 00:01:15,030
So image that competition, which is known as eyeless.

15
00:01:15,160 --> 00:01:16,920
We are see for short.

16
00:01:17,930 --> 00:01:19,410
Stand for image net.

17
00:01:19,620 --> 00:01:22,680
Large scale visual recognition challenge.

18
00:01:23,910 --> 00:01:27,360
This challenge was held between 2010 and 2017.

19
00:01:27,660 --> 00:01:35,970
Every year, participants in this challenge were given a data set of images and they had to classify

20
00:01:36,000 --> 00:01:37,920
those images into several labeled.

21
00:01:40,040 --> 00:01:47,550
In some of the competitions, the dataset had over a million observations and glasses to be identified

22
00:01:47,570 --> 00:01:48,590
but intolerant.

23
00:01:49,460 --> 00:01:51,680
So it was a very large scale competition.

24
00:01:53,930 --> 00:02:00,020
The goal of this challenge was to promote the development of better computer vision techniques.

25
00:02:01,340 --> 00:02:03,860
And also to benchmark the state of the art.

26
00:02:04,580 --> 00:02:12,610
So the winner of each year, that convolutional network was the benchmark network.

27
00:02:12,860 --> 00:02:15,440
It was considered as the best of its time.

28
00:02:18,910 --> 00:02:21,310
So here are some popular CNN architectures.

29
00:02:21,940 --> 00:02:25,510
Some of these also won the A-list VRC challenge.

30
00:02:27,760 --> 00:02:29,610
The first one here is Leonard.

31
00:02:31,270 --> 00:02:34,990
This was the oldest and most popular CNN architecture.

32
00:02:37,150 --> 00:02:39,080
It has only 60000 parameters.

33
00:02:39,370 --> 00:02:42,890
And if you look at it, it was made in nineteen ninety eight.

34
00:02:43,930 --> 00:02:49,840
So even in 1998, convolutional neural networks were dead and they were gaining popularity.

35
00:02:51,400 --> 00:03:00,370
But a major breakthrough came in 2012 when Alex Net was able to achieve very high accuracy on previous

36
00:03:00,610 --> 00:03:02,030
image classification problems.

37
00:03:04,380 --> 00:03:14,250
This raised the interest of people in CNN and in 2013, ZF Net won the eyeless we are see challenge

38
00:03:14,810 --> 00:03:16,900
and it was a convolutional network.

39
00:03:18,900 --> 00:03:22,770
It was able to achieve accuracy rate of nearly 85 percent.

40
00:03:24,600 --> 00:03:29,820
In 2014, we got two very popular architectures.

41
00:03:30,180 --> 00:03:34,020
One is Google led and the other is reading it.

42
00:03:35,880 --> 00:03:36,990
Google it was the winner.

43
00:03:37,080 --> 00:03:37,500
We did it.

44
00:03:37,500 --> 00:03:41,720
Net was runnerup Google.

45
00:03:41,730 --> 00:03:50,110
It had four million parameters as compared to Reginette, which had 138 million parameters and does

46
00:03:50,160 --> 00:03:50,730
zone five.

47
00:03:50,820 --> 00:03:55,120
The winner was arrest net and in 2016 and 17.

48
00:03:55,150 --> 00:03:57,730
Also, there were two other architectures which won.

49
00:03:58,410 --> 00:04:02,040
But these are the most popular architectures that you should know about.

50
00:04:03,510 --> 00:04:05,460
We look at Linnet.

51
00:04:06,120 --> 00:04:09,660
Google it and we in it in more detail.
