sample.en.srt

1
00:00:00,000 --> 00:00:02,430
Coding has been
the bread and butter for

2
00:00:02,430 --> 00:00:04,290
developers since
the dawn of computing.

3
00:00:04,290 --> 00:00:07,170
We're used to creating
applications by breaking down

4
00:00:07,170 --> 00:00:09,330
requirements into
composable problems

5
00:00:09,330 --> 00:00:11,145
that can then be coded against.

6
00:00:11,145 --> 00:00:13,350
So for example, if
we have to write

7
00:00:13,350 --> 00:00:15,990
an application that figures
out a stock analytic,

8
00:00:15,990 --> 00:00:18,480
maybe the price
divided by the ratio,

9
00:00:18,480 --> 00:00:20,400
we can usually write code to get

10
00:00:20,400 --> 00:00:22,380
the values from a data source,

11
00:00:22,380 --> 00:00:25,050
do the calculation and
then return the result.

12
00:00:25,050 --> 00:00:27,180
Or if we're writing a game

13
00:00:27,180 --> 00:00:29,010
we can usually figure
out the rules.

14
00:00:29,010 --> 00:00:30,690
For example, if the ball hits

15
00:00:30,690 --> 00:00:32,160
the brick then the brick should

16
00:00:32,160 --> 00:00:34,230
vanish and the ball
should rebound.

17
00:00:34,230 --> 00:00:36,180
But if the ball falls
off the bottom of

18
00:00:36,180 --> 00:00:38,610
the screen then maybe
the player loses their life.

19
00:00:38,610 --> 00:00:41,355
We can represent that
with this diagram.

20
00:00:41,355 --> 00:00:44,225
Rules and data go in
answers come out.

21
00:00:44,225 --> 00:00:45,620
Rules are expressed in

22
00:00:45,620 --> 00:00:48,320
a programming language
and data can come from

23
00:00:48,320 --> 00:00:50,090
a variety of sources from

24
00:00:50,090 --> 00:00:53,125
local variables all the
way up to databases.

25
00:00:53,125 --> 00:00:56,930
Machine learning
rearranges this diagram

26
00:00:56,930 --> 00:00:58,475
where we put answers in data

27
00:00:58,475 --> 00:01:00,605
in and then we get rules out.

28
00:01:00,605 --> 00:01:02,945
So instead of us as
developers figuring

29
00:01:02,945 --> 00:01:05,510
out the rules when should
the brick be removed,

30
00:01:05,510 --> 00:01:07,205
when should
the player's life end,

31
00:01:07,205 --> 00:01:10,310
or what's the desired analytic
for any other concept,

32
00:01:10,310 --> 00:01:12,170
what we will do is we
can get a bunch of

33
00:01:12,170 --> 00:01:13,850
examples for what we want to

34
00:01:13,850 --> 00:01:17,560
see and then have the computer
figure out the rules.

35
00:01:17,560 --> 00:01:20,360
Now, this is particularly
valuable for problems that

36
00:01:20,360 --> 00:01:23,435
you can't solve by figuring
the rules out for yourself.

37
00:01:23,435 --> 00:01:26,765
So consider this example,
activity recognition.

38
00:01:26,765 --> 00:01:29,870
If I'm building a device that
detects if somebody is say

39
00:01:29,870 --> 00:01:32,975
walking and I have data
about their speed,

40
00:01:32,975 --> 00:01:34,970
I might write code like this

41
00:01:34,970 --> 00:01:36,830
and if they're
running well that's

42
00:01:36,830 --> 00:01:39,170
a faster speed so I
could adapt my code

43
00:01:39,170 --> 00:01:41,580
to this and if they're biking,

44
00:01:41,580 --> 00:01:43,025
well that's not too bad either.

45
00:01:43,025 --> 00:01:45,260
I can adapt my code like this.

46
00:01:45,260 --> 00:01:47,780
But then I have to do
golf recognition too,

47
00:01:47,780 --> 00:01:49,825
now my concept becomes broken.

48
00:01:49,825 --> 00:01:51,080
But not only that,

49
00:01:51,080 --> 00:01:53,750
doing it by speed alone
of course is quite naive.

50
00:01:53,750 --> 00:01:56,150
We walk and run at
different speeds uphill and

51
00:01:56,150 --> 00:01:57,620
downhill and other people

52
00:01:57,620 --> 00:01:59,390
walk and run at
different speeds to us.

53
00:01:59,390 --> 00:02:01,820
So, let's go back
to this diagram.

54
00:02:01,820 --> 00:02:04,050
Ultimately machine
learning is very

55
00:02:04,050 --> 00:02:06,770
similar but we're just
flipping the axes.

56
00:02:06,770 --> 00:02:09,350
So instead of me trying
to express the problem as

57
00:02:09,350 --> 00:02:12,170
rules when often that
isn't even possible,

58
00:02:12,170 --> 00:02:13,595
I'll have to compromise.

59
00:02:13,595 --> 00:02:15,470
The new paradigm is that I get

60
00:02:15,470 --> 00:02:17,720
lots and lots of
examples and then I have

61
00:02:17,720 --> 00:02:20,150
labels on those
examples and I use

62
00:02:20,150 --> 00:02:23,105
the data to say this is
what walking looks like,

63
00:02:23,105 --> 00:02:24,350
this is what running looks like,

64
00:02:24,350 --> 00:02:26,510
this is what biking
looks like and yes,

65
00:02:26,510 --> 00:02:28,400
even this is what
golfing looks like.

66
00:02:28,400 --> 00:02:30,980
So, then it becomes
answers and data

67
00:02:30,980 --> 00:02:34,535
in with rules being
inferred by the machine.

68
00:02:34,535 --> 00:02:36,710
A machine learning algorithm then

69
00:02:36,710 --> 00:02:38,645
figures out the
specific patterns in

70
00:02:38,645 --> 00:02:40,160
each set of data that

71
00:02:40,160 --> 00:02:42,650
determines the
distinctiveness of each.

72
00:02:42,650 --> 00:02:44,270
That's what's so powerful and

73
00:02:44,270 --> 00:02:46,660
exciting about
this programming paradigm.

74
00:02:46,660 --> 00:02:48,440
It's more than just a new way

75
00:02:48,440 --> 00:02:49,775
of doing the same old thing.

76
00:02:49,775 --> 00:02:51,770
It opens up new possibilities

77
00:02:51,770 --> 00:02:53,905
that were infeasible
to do before.

78
00:02:53,905 --> 00:02:55,270
So in the next few minutes,

79
00:02:55,270 --> 00:02:57,680
I'm going to show you
the basics of creating

80
00:02:57,680 --> 00:02:59,360
a neural network which is

81
00:02:59,360 --> 00:03:02,435
the workhorse of doing this type
of pattern recognition.

82
00:03:02,435 --> 00:03:04,040
A neural network is

83
00:03:04,040 --> 00:03:06,200
just a slightly more
advanced implementation of

84
00:03:06,200 --> 00:03:08,695
machine learning and we
call that deep learning.

85
00:03:08,695 --> 00:03:11,330
But fortunately it's
actually very easy to code.

86
00:03:11,330 --> 00:03:13,430
So, we're just going to jump
straight into deep learning.

87
00:03:13,430 --> 00:03:16,130
We'll start with a simple one
and then we'll move on

88
00:03:16,130 --> 00:03:17,675
to one that does computer vision

89
00:03:17,675 --> 00:03:19,250
in about 10 lines of code.

90
00:03:19,250 --> 00:03:22,085
But let's start with a very
simple "Hello World" example.

91
00:03:22,085 --> 00:03:24,870
So you can see just how
everything hangs together.