1
00:00:00,150 --> 00:00:06,300
Hi and welcome to our lecture on generative adversarial neural networks, or Gans for short.

2
00:00:06,780 --> 00:00:07,890
So let's get started.

3
00:00:08,970 --> 00:00:10,770
So what's in store in this section?

4
00:00:10,800 --> 00:00:15,210
Well, firstly, I'll tell you what guns are and why they're so exciting.

5
00:00:15,510 --> 00:00:20,130
I'll give you some examples of guns that have been put out there by researchers over the years.

6
00:00:20,640 --> 00:00:26,010
I'll tell you how guns work and then how you can go about treating your own guns, as well as some of

7
00:00:26,010 --> 00:00:27,660
the challenges you will experience.

8
00:00:28,080 --> 00:00:34,110
And then I can give you some practical implementations of guns because they may not be immediately obvious

9
00:00:34,110 --> 00:00:34,500
to you.

10
00:00:34,770 --> 00:00:36,210
Why guns are so useful.

11
00:00:36,780 --> 00:00:40,020
And here we have some images of some gun outputs.

12
00:00:40,060 --> 00:00:41,730
You can see Google's big gun.

13
00:00:42,210 --> 00:00:48,630
This is this one is actually generating fake images here that look so real, as well as first generation

14
00:00:48,630 --> 00:00:49,830
performed by Guns Hill.

15
00:00:50,310 --> 00:00:53,850
You can see all of these faces A.I. generated faces.

16
00:00:54,240 --> 00:00:55,510
They're not real people.

17
00:00:55,530 --> 00:00:57,450
So it's quite amazing, isn't it?

18
00:00:58,920 --> 00:01:01,290
So let's talk about what are guns?

19
00:01:01,650 --> 00:01:05,700
So guns were first introduced in 2014 by Ian Goodfellow.

20
00:01:06,180 --> 00:01:13,380
Guns are a type of neural network that what they do, they generate data that could have plausibly come

21
00:01:13,380 --> 00:01:15,750
from an existing distribution of symbols.

22
00:01:16,260 --> 00:01:18,360
So let's take a look at what that means.

23
00:01:18,840 --> 00:01:24,660
Here we have a sample of the endless dataset here, and in this column, we have data that has been

24
00:01:24,660 --> 00:01:28,290
generated by a gun that looks like it came from the embassy, they said.

25
00:01:28,290 --> 00:01:29,430
But it wasn't there.

26
00:01:29,850 --> 00:01:33,240
This has been artificially generated by a gun.

27
00:01:34,140 --> 00:01:35,310
Same old faces.

28
00:01:35,730 --> 00:01:38,790
Same with some images from the sapphire dataset as well.

29
00:01:39,330 --> 00:01:44,580
And for a technology, council actually said guns were one of the best innovations of the last three

30
00:01:44,580 --> 00:01:44,930
years.

31
00:01:44,940 --> 00:01:46,380
This was in 2018.

32
00:01:46,890 --> 00:01:53,250
So guns are quite exciting and quite promising, so we can take a look at some of the examples of guns

33
00:01:53,250 --> 00:01:53,910
over the years.

34
00:01:54,330 --> 00:01:56,610
This one came out in 2018.

35
00:01:56,910 --> 00:02:02,670
It's called Big Gun, and it was able to generate all of these artificially generated images here.

36
00:02:03,090 --> 00:02:03,750
Very cool.

37
00:02:04,860 --> 00:02:09,270
We can even do anime characters, which is a project will be doing in discourse later on.

38
00:02:09,810 --> 00:02:11,670
Generate some anime characters.

39
00:02:11,670 --> 00:02:17,190
So we give it a dataset of existing anime characters and we start generating new ones out of it.

40
00:02:18,600 --> 00:02:23,910
Next, we can look at something called image to image translation, where we have an input image here.

41
00:02:24,480 --> 00:02:30,030
This is the ground truth as well here for a night scene, and it generates this other dusk looking scene

42
00:02:30,030 --> 00:02:31,980
here, which is an image of these two.

43
00:02:32,460 --> 00:02:38,880
Similar to this one, you can basically ticket image that's daytime and then generate a night vision

44
00:02:38,880 --> 00:02:41,010
of that image based on this type of data set.

45
00:02:42,000 --> 00:02:48,240
Also, this one called sketches two images, so you can just feed again a sketch like this, and it

46
00:02:48,240 --> 00:02:53,680
creates an output like this so they can create a person, can create a backpack, you can create a show,

47
00:02:53,730 --> 00:02:54,900
as we can see here.

48
00:02:55,650 --> 00:02:56,760
Pretty awesome, isn't it?

49
00:02:57,720 --> 00:03:02,720
Next, we can text to image translations, so give it some ticks.

50
00:03:02,730 --> 00:03:09,750
The small bird has a redhead with feathers that feed from red to gray, from head to tail, and it generates

51
00:03:09,750 --> 00:03:12,570
images of a bird that fits that description.

52
00:03:13,140 --> 00:03:17,880
That is extremely impressive, in my opinion, and this is another example of it here.

53
00:03:19,290 --> 00:03:22,200
Also, we have semantic image to photo translation.

54
00:03:22,650 --> 00:03:28,530
So later on in discourse, you will look at things that segmentation models that do segmentation for

55
00:03:28,530 --> 00:03:34,500
a scene so we can separate to cause him to ruin and the people and trees as built in buildings as well.

56
00:03:34,920 --> 00:03:38,360
Well, what if we were to tweet you were to take this image here?

57
00:03:38,400 --> 00:03:43,230
This segmented image here and generate photorealistic scene out of it?

58
00:03:43,770 --> 00:03:45,210
That's what these scans do.

59
00:03:45,300 --> 00:03:53,430
The reverse engineer, that's semantic representation and very popular one which I have used in real

60
00:03:53,430 --> 00:03:56,820
life production scenarios many times.

61
00:03:57,300 --> 00:03:58,100
That's cool.

62
00:03:58,110 --> 00:03:59,580
That's super resolution, Garneau.

63
00:03:59,700 --> 00:04:03,780
So again, for short, you can take a lower resolution image here.

64
00:04:03,810 --> 00:04:10,320
So what if you were using a CCTV camera that was getting a blurry feed of see a license plate you can

65
00:04:10,320 --> 00:04:16,890
use as Horrigan to enhance that image and then pass it to you will see our and it probably will improve

66
00:04:17,280 --> 00:04:19,980
by a few percentage points in accuracy.

67
00:04:20,280 --> 00:04:22,080
That's what I've experienced, in my opinion.

68
00:04:24,120 --> 00:04:29,670
So next, we take a look at how guns work, so stay tuned for that lesson.

69
00:04:29,820 --> 00:04:30,690
Thank you for watching.