1
00:00:00,000 --> 00:00:01,870
So, there we saw
that the classifier

2
00:00:01,870 --> 00:00:03,880
that we built for
classifying fashion,

3
00:00:03,880 --> 00:00:05,680
like using convolutions
were able to

4
00:00:05,680 --> 00:00:07,870
make it more efficient and
to make it more accurate.

5
00:00:07,870 --> 00:00:09,400
I thought that was
really, really cool,

6
00:00:09,400 --> 00:00:10,900
but it's still very limited in

7
00:00:10,900 --> 00:00:12,520
the scenario because
all of our images

8
00:00:12,520 --> 00:00:15,820
are 28 by 28 and the subject
is actually centered.

9
00:00:15,820 --> 00:00:18,070
And while it's a fabulous
dataset for learning,

10
00:00:18,070 --> 00:00:19,270
it's like when we start getting

11
00:00:19,270 --> 00:00:20,470
into real-world images and

12
00:00:20,470 --> 00:00:21,940
complex images that maybe we

13
00:00:21,940 --> 00:00:23,870
need to go a little bit
further. What do you think?

14
00:00:23,870 --> 00:00:26,755
I think it's really cool
that taking the core idea

15
00:00:26,755 --> 00:00:29,320
of a confinet allows you

16
00:00:29,320 --> 00:00:31,930
to implement
an algorithm to confine

17
00:00:31,930 --> 00:00:33,280
not just handbags right in

18
00:00:33,280 --> 00:00:35,050
the middle of the image
but anywhere in the image,

19
00:00:35,050 --> 00:00:36,850
so it could be carried by someone

20
00:00:36,850 --> 00:00:38,770
on the left or the right
of a much bigger and,

21
00:00:38,770 --> 00:00:40,120
say, a one-megapixel image.

22
00:00:40,120 --> 00:00:42,340
This is 1000 by 1000 pixels.

23
00:00:42,340 --> 00:00:46,625
Also for many applications
rather than using grayscale,

24
00:00:46,625 --> 00:00:48,050
want to use color images-

25
00:00:48,050 --> 00:00:48,350
All right.

26
00:00:48,350 --> 00:00:50,970
And the same core ideas
but with a bigger dataset,

27
00:00:50,970 --> 00:00:54,010
bigger images in similar labels
lets you do that.

28
00:00:54,010 --> 00:00:55,580
All right. So, the technique

29
00:00:55,580 --> 00:00:57,200
that you're learning in this,

30
00:00:57,200 --> 00:00:59,000
is really really
helping you to be able

31
00:00:59,000 --> 00:01:01,910
to succeed in these more
real-world scenarios.

32
00:01:01,910 --> 00:01:05,205
So, I know you've been working
on a dataset on horses-

33
00:01:05,205 --> 00:01:05,385
Yeah.

34
00:01:05,385 --> 00:01:06,165
And humans.

35
00:01:06,165 --> 00:01:07,760
Yeah, that's been a lot of fun.

36
00:01:07,760 --> 00:01:09,250
I've been working on a dataset

37
00:01:09,250 --> 00:01:10,520
that's a number of images of

38
00:01:10,520 --> 00:01:12,320
horses and they're moving

39
00:01:12,320 --> 00:01:14,180
around the image and
they're in different poses,

40
00:01:14,180 --> 00:01:15,845
and humans in the same way

41
00:01:15,845 --> 00:01:17,870
and diverse humans male, female,

42
00:01:17,870 --> 00:01:18,890
different races, that kind of

43
00:01:18,890 --> 00:01:20,030
thing to see if we can build

44
00:01:20,030 --> 00:01:22,640
a binary classifier
between the two of them.

45
00:01:22,640 --> 00:01:24,290
But what was really
interesting about this is

46
00:01:24,290 --> 00:01:26,000
that they're all
computer-generated images,

47
00:01:26,000 --> 00:01:27,170
but we can use them to classify

48
00:01:27,170 --> 00:01:29,220
real photos. I had
a lot of fun with that.

49
00:01:29,220 --> 00:01:31,220
So, I think there'll
be a fun exercise

50
00:01:31,220 --> 00:01:33,035
for you to work on as well.

51
00:01:33,035 --> 00:01:34,760
And if you're ever wondering of

52
00:01:34,760 --> 00:01:36,170
these algorithms you're
learning whether

53
00:01:36,170 --> 00:01:37,730
this is the real stuff,

54
00:01:37,730 --> 00:01:40,340
the algorithims you're learning
is really the real stuff

55
00:01:40,340 --> 00:01:42,890
that is used today in
many commercial applications.

56
00:01:42,890 --> 00:01:44,360
For example, if you look at

57
00:01:44,360 --> 00:01:46,610
the way a real
self-driving car today

58
00:01:46,610 --> 00:01:48,425
uses cameras to detect

59
00:01:48,425 --> 00:01:50,990
other vehicles or pedestrians
to try to avoid them,

60
00:01:50,990 --> 00:01:53,090
they use convolutional
neural networks

61
00:01:53,090 --> 00:01:54,529
for that part of the task,

62
00:01:54,529 --> 00:01:56,180
very similar to what
you are learning.

63
00:01:56,180 --> 00:01:58,310
And in fact, in other contexts,

64
00:01:58,310 --> 00:01:59,400
I've heard you speak about

65
00:01:59,400 --> 00:02:01,110
using a convolutional
neural network.

66
00:02:01,110 --> 00:02:02,890
To take a picture, for example.

67
00:02:02,890 --> 00:02:05,270
Yeah, we can take a picture
of a crop and try to tell

68
00:02:05,270 --> 00:02:07,790
if it has a disease coming.
So, that was really cool.

69
00:02:07,790 --> 00:02:09,895
Oh, thank you, thank
you. That's really fun.

70
00:02:09,895 --> 00:02:11,725
So, in the next video,

71
00:02:11,725 --> 00:02:14,390
you'll learn how to apply
convolutional neural networks

72
00:02:14,390 --> 00:02:17,390
to these much bigger and
more complex images.

73
00:02:17,390 --> 00:02:19,860
Please go on to the next video.