Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated:
1
00:00:00,420 --> 00:00:05,070
Hello and welcome back to the course on computer vision in today's tutorial We'll talk about the idea
2
00:00:05,070 --> 00:00:06,670
behind Gad's.
3
00:00:06,670 --> 00:00:14,550
Guns were invented by a good fellow or to be entirely correct by Dr good fellow who currently works
4
00:00:14,610 --> 00:00:20,960
at openly as a researcher and he's only 31 years old.
5
00:00:20,970 --> 00:00:27,030
He actually created Gannes even earlier so when he was in in his late 20s a couple of years ago he created
6
00:00:27,030 --> 00:00:34,950
Gans and the story goes that he finished off the idea of guns and when he was at a bar none of that's
7
00:00:34,950 --> 00:00:39,120
true or not but you can definitely check out more of that online.
8
00:00:39,120 --> 00:00:41,690
But it's a crazy story anyway.
9
00:00:42,450 --> 00:00:50,430
How such inside such a young age has brought something to the world that is really drastically changing
10
00:00:50,430 --> 00:00:55,460
this whole field of neural networks and computer vision.
11
00:00:55,560 --> 00:01:02,370
And so what are Gad's the creation grandstands for generative adversarial network.
12
00:01:02,550 --> 00:01:09,840
And we're going to talk more about what Gannes are and how they're shocked at how they're trained what
13
00:01:09,840 --> 00:01:12,640
they're used for in the further trolls in this section.
14
00:01:12,670 --> 00:01:15,080
But I just want to talk about the philosophy.
15
00:01:15,090 --> 00:01:18,110
Why were they created in the first place.
16
00:01:18,360 --> 00:01:27,660
Well the reason why Gannes were constructed is because neural networks are generally good at predicting
17
00:01:27,660 --> 00:01:31,770
things or classifying things and solving problems.
18
00:01:31,800 --> 00:01:41,640
But we wanted to create or I guess to say that researchers AI researchers wanted to create a type of
19
00:01:41,640 --> 00:01:48,220
neural network that we can create for itself and that's why they came up with this idea of generative
20
00:01:48,240 --> 00:01:49,980
adversarial networks.
21
00:01:49,980 --> 00:02:00,090
So Gans can actually generate images they can create images of all kinds of photos of things that never
22
00:02:00,090 --> 00:02:01,980
actually ever existed before.
23
00:02:02,130 --> 00:02:11,620
So they learn about the world they learn about our objects that we have that we use or animals or anything.
24
00:02:11,790 --> 00:02:18,150
And then they can create new versions of those objects that actually never walk this planet or never
25
00:02:18,600 --> 00:02:24,600
these objects never existed were never built by humans but they can create images of those objects so
26
00:02:24,600 --> 00:02:29,010
they can actually if you see you can think about it as they can have an imagination.
27
00:02:29,160 --> 00:02:33,330
And the way that Gannes work is they have two components.
28
00:02:33,330 --> 00:02:36,490
They have a generator and a discriminator.
29
00:02:37,080 --> 00:02:39,960
And these two components are constantly in touch.
30
00:02:40,050 --> 00:02:47,880
The generator generates images and the discriminator then assesses those images and tells the generator
31
00:02:47,880 --> 00:02:56,670
whether or not those images are likely to be similar or are similar to what it has actually seen in
32
00:02:56,670 --> 00:02:59,850
the real world that already exists.
33
00:02:59,850 --> 00:03:02,920
So it's got these two components at work in tandem.
34
00:03:03,060 --> 00:03:05,790
And when you're training this is the interesting part.
35
00:03:05,780 --> 00:03:10,470
This is like a very philosophical part really that will definitely go into the detail about how the
36
00:03:10,470 --> 00:03:13,620
generator works how it works how they connect and so on.
37
00:03:13,620 --> 00:03:20,550
But what we need to understand is the philosophical concept that when you're creating a degenerative
38
00:03:20,640 --> 00:03:28,260
adversarial network when you're training up you are bringing these two up together from scratch so it's
39
00:03:28,260 --> 00:03:29,340
not like you have.
40
00:03:29,400 --> 00:03:35,490
And this is the part where I needed to get my head around when I was learning about them is that you're
41
00:03:35,490 --> 00:03:40,890
not just creating is it's like you have a discriminator that already knows everything about the world.
42
00:03:40,980 --> 00:03:47,730
And then it is this generator is generating things and the discriminator can't right away tell the generator
43
00:03:47,790 --> 00:03:50,710
if that is you know that looks like a real object or not.
44
00:03:50,850 --> 00:03:56,740
No the fact is that they start from scratch from zero and then they learn together.
45
00:03:56,850 --> 00:04:04,860
So the generator generate some images for instance you're generating images of tables and then the discriminator
46
00:04:04,860 --> 00:04:09,580
will look at the images that the generated generated and will look at some images of real tables to
47
00:04:09,630 --> 00:04:16,760
compare the two and it will learn for itself what is a table what isn't a table because it all know
48
00:04:16,770 --> 00:04:22,710
but by default it will know what the generator generates actually doesn't exist in it.
49
00:04:23,010 --> 00:04:29,180
It has that knowledge and therefore it will know what is a table what isn't a table and then it will
50
00:04:29,180 --> 00:04:35,220
also give feedback to the generator saying hey your tables are not really tables because of this and
51
00:04:35,220 --> 00:04:35,730
this and this.
52
00:04:35,730 --> 00:04:36,680
I could tell right away.
53
00:04:36,690 --> 00:04:38,430
So they were kind of playing a game there.
54
00:04:38,490 --> 00:04:41,610
Think of it as are against each other but they're learning together.
55
00:04:41,610 --> 00:04:47,640
So the general is always trying to create images which look like real tables to fool the discriminator
56
00:04:47,640 --> 00:04:52,530
to trick the discriminator in believing that those are real tables whereas the real discrimination always
57
00:04:52,530 --> 00:04:59,780
has actual real to images of actual real tables to compare to in order to learn better.
58
00:04:59,910 --> 00:05:02,340
And then also give feedback to the generator.
59
00:05:02,400 --> 00:05:06,600
And in that sense and we'll see exactly the mechanics of this for it in that sense they learn together
60
00:05:06,600 --> 00:05:17,190
they grow together and eventually this whole network as a whole it learns how to better and better and
61
00:05:17,180 --> 00:05:25,140
better create these tables that are indistinguishable from the real world tables and zone even though
62
00:05:25,160 --> 00:05:30,240
they're working against each other fighting against each other as a whole the network is getting better
63
00:05:30,270 --> 00:05:32,540
and better and better and more robust.
64
00:05:32,580 --> 00:05:43,740
So that's the philosophy behind Gannes that they are designed to create objects or images to be to be
65
00:05:43,980 --> 00:05:51,330
exactly and be entirely accurate images of objects animals or even humans based on what we have in the
66
00:05:51,330 --> 00:05:51,710
real world.
67
00:05:51,710 --> 00:05:58,050
So you can think of it as a general adversarial network who's actually learning about our world and
68
00:05:58,050 --> 00:06:07,680
then creating objects or images of objects which actually are very similar to what we have in our world
69
00:06:07,710 --> 00:06:09,500
but never actually existed.
70
00:06:10,050 --> 00:06:11,400
So yeah that's us.
71
00:06:11,460 --> 00:06:12,510
I hope that sounds exciting.
72
00:06:12,510 --> 00:06:14,780
We're going to delve into that in the upcoming tutorials.
73
00:06:14,880 --> 00:06:21,430
And for now if you'd like some additional reading there's a great blog post by Chan chan SORN.
74
00:06:21,600 --> 00:06:22,550
So torn.
75
00:06:22,590 --> 00:06:24,720
I hope I pronounced that right.
76
00:06:25,290 --> 00:06:27,690
It's on Hakkar noon and you can check it out there.
77
00:06:27,720 --> 00:06:28,800
I've got a link.
78
00:06:28,800 --> 00:06:33,360
Basically it will talk you through the intuition behind Ganns.
79
00:06:33,490 --> 00:06:39,840
Well we're going to discuss all of this in upcoming tutorials but in case you wanted to prepare in case
80
00:06:39,840 --> 00:06:45,720
you wanted to do some pre-reading so that your you know you have an additional point of view then you
81
00:06:45,720 --> 00:06:49,260
can check out some extra insights over there.
82
00:06:49,380 --> 00:06:52,730
But again we will discuss all of this in upcoming tutorials.
83
00:06:52,920 --> 00:06:54,410
And I can't wait to see you there.
84
00:06:54,570 --> 00:06:55,450
Until next time.
85
00:06:55,620 --> 00:06:56,660
Enjoy computer vision.
9634
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.