1
00:00:01,170 --> 00:00:07,080
Welcome back to the course in this section, we'll take a look at the grandfather of CNN's and that's

2
00:00:07,080 --> 00:00:09,750
the line at CNN architecture.

3
00:00:10,320 --> 00:00:11,250
So let's get started.

4
00:00:11,370 --> 00:00:13,940
So firstly, why was Lynnette developed?

5
00:00:13,980 --> 00:00:16,860
Well, the US Postal Service had a problem.

6
00:00:17,310 --> 00:00:22,050
They had a problem in automatically reading these digits, these handwritten digits when people write

7
00:00:22,050 --> 00:00:25,050
their mailing address on the letters or packages.

8
00:00:25,650 --> 00:00:32,130
They needed a way to read these numbers accurately, and that was the whole genesis of why that was

9
00:00:32,130 --> 00:00:33,390
developed for this problem.

10
00:00:33,840 --> 00:00:40,950
It was actually first introduced in 1995, and it got published and publicly announced an open sourced

11
00:00:41,310 --> 00:00:42,890
in nineteen nine, I believe.

12
00:00:42,900 --> 00:00:46,940
But the initial CNN was developed in 1995 by Liqun.

13
00:00:47,820 --> 00:00:53,610
It was developed through handwritten digits, and you may have known this by now, but the amnesty to

14
00:00:53,610 --> 00:00:57,900
said was basically collected from the U.S. Postal Service data.

15
00:00:58,500 --> 00:01:03,300
So all those experiments we were doing with Amnesty, the set that was basically collected from the

16
00:01:03,300 --> 00:01:09,660
first well on the first CNN, obviously, but it was collected from one of the first big data sets that

17
00:01:09,660 --> 00:01:12,390
were used in computer vision and neural networks.

18
00:01:12,960 --> 00:01:13,680
Pretty cool, huh?

19
00:01:14,940 --> 00:01:17,280
So now let's take a look at a little architecture.

20
00:01:17,730 --> 00:01:19,050
So this should look familiar.

21
00:01:19,680 --> 00:01:21,090
Here we have an input image.

22
00:01:21,150 --> 00:01:23,550
We have six feature maps.

23
00:01:24,030 --> 00:01:28,550
We have a filter size that is actually whereas five by five.

24
00:01:28,560 --> 00:01:33,150
So not that big, not that small, because we've used too much and filters previously.

25
00:01:33,870 --> 00:01:35,040
Then we have a max pool.

26
00:01:35,040 --> 00:01:40,290
If you can see the dimensions shrink by two, so have the six feature maps here.

27
00:01:40,830 --> 00:01:48,780
Next, we have 16 filters here filters and these are file size five by five.

28
00:01:49,350 --> 00:01:54,180
This is the output size here, and then we shrink it by having a next max pool.

29
00:01:54,180 --> 00:01:56,100
In a max boot is a boot max.

30
00:01:56,100 --> 00:02:01,350
Players have to buy two kernels strike two and then we have to fully connected layers.

31
00:02:01,350 --> 00:02:03,420
So we actually have two fully connected layers.

32
00:02:03,840 --> 00:02:11,170
We have 120 units here where neurons and then 80 for neurons here before finally it goes to the output

33
00:02:11,190 --> 00:02:11,640
here.

34
00:02:12,330 --> 00:02:15,450
This is basically a description of that same architecture.

35
00:02:15,510 --> 00:02:16,440
So I wouldn't read it out.

36
00:02:16,950 --> 00:02:23,910
But just so you know, this simple architecture with just six filters here and 16 filters here was able

37
00:02:23,910 --> 00:02:27,440
to get ninety nine point three percent accuracy on the amnesty, was it?

38
00:02:27,930 --> 00:02:28,830
And that's pretty cool.

39
00:02:29,430 --> 00:02:30,480
Now let's take a look at this.

40
00:02:30,840 --> 00:02:34,550
This is an actual demo of the Linnet architecture.

41
00:02:34,560 --> 00:02:40,320
Back then, you can see as numbers slide across here, it's been classified on top here, so you can

42
00:02:40,320 --> 00:02:40,920
see what they are.

43
00:02:43,130 --> 00:02:45,050
So this is a pretty cool demonstration, isn't it?

44
00:02:46,130 --> 00:02:51,470
So we'll stop there for now because that's basically all there is to to then.

45
00:02:51,470 --> 00:02:56,240
That's not a very complicated network, as you've seen throughout a simple it's rather similar to the

46
00:02:56,240 --> 00:02:57,680
ones we've done before, actually.

47
00:02:58,300 --> 00:03:00,780
Now we'll take a look at it.

48
00:03:01,280 --> 00:03:03,350
So I'll see you in the next section.

49
00:03:03,530 --> 00:03:04,010
Thank you.