1
00:00:00,000 --> 00:00:04,275
[MUSIC]

2
00:00:04,275 --> 00:00:07,910
Okay, given your great historical
perspective of machine learning-

3
00:00:07,910 --> 00:00:08,680
>> Are you calling me old?

4
00:00:08,680 --> 00:00:10,010
>> I am calling you old.

5
00:00:10,010 --> 00:00:12,075
>> It's always like this.

6
00:00:12,075 --> 00:00:16,566
>> [LAUGH] I was wondering if you
could give me a little perspective on

7
00:00:16,566 --> 00:00:17,488
the future.

8
00:00:17,488 --> 00:00:18,951
>> Well, let's talk about the present.

9
00:00:18,951 --> 00:00:19,784
>> Okay.

10
00:00:19,784 --> 00:00:23,250
>> [LAUGH]
>> I'll accept that,

11
00:00:23,250 --> 00:00:24,870
we can talk about the present.

12
00:00:24,870 --> 00:00:27,890
He always makes things really long and
really detailed.

13
00:00:27,890 --> 00:00:29,825
So just talk about the present.

14
00:00:29,825 --> 00:00:31,714
We'll get to the future eventually,
don't worry.

15
00:00:31,714 --> 00:00:36,620
>> [LAUGH]
>> Here we go, the present.

16
00:00:36,620 --> 00:00:38,270
What is the state of machine learning,
Carlos?

17
00:00:38,270 --> 00:00:41,934
>> [CROSSTALK] I'm really excited about
machine learning today, honestly.

18
00:00:41,934 --> 00:00:44,560
Being old, that you work in
machine learning for a long time.

19
00:00:44,560 --> 00:00:48,590
But today, every time I open a web app,
it uses machine learning inside it.

20
00:00:48,590 --> 00:00:50,346
And that's pretty cool.

21
00:00:50,346 --> 00:00:53,109
And I pull out my phone, where's my phone?

22
00:00:53,109 --> 00:00:54,552
You always have your phone with you.

23
00:00:54,552 --> 00:00:56,482
>> I actually don't have
my phone right now.

24
00:00:56,482 --> 00:00:57,450
>> She's always on her phone.

25
00:00:58,800 --> 00:01:02,350
Anyway, when you pull out your phone,
every time you use it, there's

26
00:01:02,350 --> 00:01:06,550
a system that uses machine learning
right inside the cloud that supports it.

27
00:01:06,550 --> 00:01:08,164
Which is really pretty cool.

28
00:01:08,164 --> 00:01:10,820
And now with devices, check this out.

29
00:01:10,820 --> 00:01:12,525
I got this watch.

30
00:01:12,525 --> 00:01:15,300
>> Pretty spiffy.
>> It's a wonderful watch that

31
00:01:15,300 --> 00:01:18,680
tracks my activity,
knows when I'm running, when I'm biking-

32
00:01:18,680 --> 00:01:19,680
>> Knows when he's sleeping or

33
00:01:19,680 --> 00:01:20,360
not sleeping.

34
00:01:20,360 --> 00:01:23,192
Always this debate,
who got less sleep with the baby?

35
00:01:23,192 --> 00:01:24,619
[LAUGH]
>> [LAUGH] So, yeah,

36
00:01:24,619 --> 00:01:26,344
that's a wonderful watch.

37
00:01:26,344 --> 00:01:29,140
And underneath these things,
it's all using machine learning.

38
00:01:29,140 --> 00:01:32,156
So we see the power of machine
learning just kind of, again and

39
00:01:32,156 --> 00:01:33,780
again, being used today.

40
00:01:33,780 --> 00:01:37,770
But I think we've only scratched the
surface of what machine learning can do.

41
00:01:37,770 --> 00:01:40,320
Where do you see machine
learning going in the future?

42
00:01:40,320 --> 00:01:40,902
You tell me.

43
00:01:40,902 --> 00:01:42,323
>> Where do I see it?

44
00:01:42,323 --> 00:01:47,045
Well, I know probably a lot of ways in
which it would be useful to Carlos.

45
00:01:47,045 --> 00:01:49,935
One being the fact that he hates driving.

46
00:01:49,935 --> 00:01:56,455
So of course, there is this smart car idea
that this self-driving, not smart car.

47
00:01:56,455 --> 00:01:57,635
>> They're also smart.

48
00:01:57,635 --> 00:01:58,775
>> Are they smart?

49
00:01:58,775 --> 00:02:00,055
They're pretty smart, I guess.

50
00:02:00,055 --> 00:02:01,655
It's a self-driving car.

51
00:02:01,655 --> 00:02:02,785
A smart car is a brand.

52
00:02:02,785 --> 00:02:04,965
That's not what I mean,
I don't mean those little cars.

53
00:02:04,965 --> 00:02:09,850
I mean normal cars [LAUGH]
that drive themselves.

54
00:02:09,850 --> 00:02:12,560
>> That would be really useful for
me, because I don't like driving.

55
00:02:12,560 --> 00:02:13,920
I only like biking.

56
00:02:13,920 --> 00:02:16,575
And so I need some-
>> Self-driving bikes.

57
00:02:16,575 --> 00:02:17,260
>> No, no, I like driving my bike.

58
00:02:17,260 --> 00:02:19,290
>> [CROSSTALK] That's for
when you get really, really old.

59
00:02:19,290 --> 00:02:20,946
[LAUGH]
>> [CROSSTALK] So

60
00:02:20,946 --> 00:02:22,312
I'm gonna talk about that.

61
00:02:22,312 --> 00:02:26,317
But when I think about me getting old,
I think about how machine learning can

62
00:02:26,317 --> 00:02:29,670
really have impact with
your personalized medicine.

63
00:02:29,670 --> 00:02:32,550
And the question here is,
why is it when I go to a doctor for

64
00:02:32,550 --> 00:02:34,310
certain disease or certain condition,

65
00:02:34,310 --> 00:02:37,870
the treatment that Emily gets is exactly
the same as the treatment that I get.

66
00:02:37,870 --> 00:02:41,315
I mean, we're different people
with different lifestyles,

67
00:02:41,315 --> 00:02:43,720
with different body types, as you can see.

68
00:02:43,720 --> 00:02:47,160
And still, with medicine,
we're getting exactly the same treatment.

69
00:02:47,160 --> 00:02:48,260
That makes no sense.

70
00:02:48,260 --> 00:02:51,780
The treatment needs to be
personalized to who I am, what I do,

71
00:02:51,780 --> 00:02:55,800
how much of a couch potato I am,
what my DNA looks like.

72
00:02:55,800 --> 00:02:58,430
And with machine learning, we're going
to be able to do that all the time.

73
00:02:58,430 --> 00:02:59,950
It's really super cool,
I'm pretty excited.

74
00:02:59,950 --> 00:03:02,770
>> That was just so
optimistic, which is good.

75
00:03:02,770 --> 00:03:04,570
It's really nice, but it's hard!

76
00:03:04,570 --> 00:03:06,308
It's really, really hard to do this.

77
00:03:06,308 --> 00:03:11,850
The data is really noisy, cuz there's some
doctor who's there taking these records,

78
00:03:11,850 --> 00:03:15,770
making notes about Carlos' health
conditions compared to mine.

79
00:03:15,770 --> 00:03:18,370
And then there's lots of missing data,
because

80
00:03:18,370 --> 00:03:22,490
maybe Carlos skipped his last checkup or
the doctor forgot to enter something.

81
00:03:22,490 --> 00:03:26,600
Or maybe the doctor put something under
a different entry than they did for me.

82
00:03:26,600 --> 00:03:30,000
And so somehow our machine learning
algorithms are gonna have to be able to

83
00:03:30,000 --> 00:03:31,900
parse this really complex and

84
00:03:31,900 --> 00:03:36,340
noisy data, in order to do this
type of personalized medicine.

85
00:03:36,340 --> 00:03:38,020
Which I think is a possibility.

86
00:03:38,020 --> 00:03:42,280
I think it's really cool, but-
>> It is hard.

87
00:03:42,280 --> 00:03:44,840
All these other things we're
talking about, from self-driving

88
00:03:44,840 --> 00:03:48,340
cars which we're seeing as a reality
today, to personalized medicine.

89
00:03:48,340 --> 00:03:49,770
It's all hard stuff.

90
00:03:49,770 --> 00:03:53,895
But in the future, and
this is the optimistic in me,

91
00:03:53,895 --> 00:03:56,850
the world is going to
be personalized to me.

92
00:03:56,850 --> 00:03:58,715
>> [LAUGH]
>> You know what I mean?

93
00:03:58,715 --> 00:04:01,416
>> The grand goal of everything for
Carlos.

94
00:04:01,416 --> 00:04:04,247
[LAUGH]
>> [LAUGH] For me and for Emily, too.

95
00:04:04,247 --> 00:04:05,765
She can have her own world.

96
00:04:05,765 --> 00:04:11,145
So the idea that today
when we go to the Web and

97
00:04:11,145 --> 00:04:16,225
we buy things, we have to go through
very similar experiences as others.

98
00:04:16,225 --> 00:04:20,600
But as the world becomes smarter, as more
data is collected, as we come up with this

99
00:04:20,600 --> 00:04:23,670
algorithms which are robust to all
this kind of noise in the data,

100
00:04:23,670 --> 00:04:26,860
we're gonna have experiences that
are really unique to us, idiosyncratic.

101
00:04:26,860 --> 00:04:29,050
And we're not gonna suffer
as much from technology.

102
00:04:29,050 --> 00:04:31,970
Technology's gonna really adapt to us,
which is really super cool.

103
00:04:33,060 --> 00:04:35,700
>> I agree, I think that is a really,
really exciting goal.

104
00:04:35,700 --> 00:04:39,240
And I think that is one of the directions
that machine learning is going in.

105
00:04:39,240 --> 00:04:42,840
But I think something else that's really
cool is even doing the things we can

106
00:04:42,840 --> 00:04:44,000
do right now.

107
00:04:44,000 --> 00:04:49,040
But doing them as the data's coming in,
in real time, with really complicated

108
00:04:49,040 --> 00:04:53,170
different types of data sets, and
putting all this information together.

109
00:04:53,170 --> 00:04:56,710
And maybe it's just even
still recommending products,

110
00:04:56,710 --> 00:05:00,820
something we've done before in the past,
but with different types of data.

111
00:05:00,820 --> 00:05:04,380
And again, in this really fast,
rapid, online type of setting, and

112
00:05:04,380 --> 00:05:10,940
getting feedback and incorporating this
to make these types of prediction, yeah.

113
00:05:10,940 --> 00:05:14,693
>> Yeah, dealing with temporal data,
which is the topic here,

114
00:05:14,693 --> 00:05:17,956
as well as the spatial
data like location data.

115
00:05:17,956 --> 00:05:22,550
All this different kind of data sets,
images, video, audio, sensing data,

116
00:05:22,550 --> 00:05:27,504
combining all these different data sources
in real time is one of the big challenges.

117
00:05:27,504 --> 00:05:31,490
And what I'm excited to see is that
there's a lot more machine learning work

118
00:05:31,490 --> 00:05:35,602
and a lot more algorithms being developed
that try to address this confusion of

119
00:05:35,602 --> 00:05:36,890
different data types.

120
00:05:36,890 --> 00:05:39,692
And that's a really cool
topic to work on right now.

121
00:05:39,692 --> 00:05:41,399
>> For sure, so.

122
00:05:41,399 --> 00:05:43,170
>> Of course.

123
00:05:43,170 --> 00:05:47,110
>> We hope that's really exciting to you,
in addition to us.

124
00:05:47,110 --> 00:05:48,840
>> Yeah, it is exciting.

125
00:05:48,840 --> 00:05:53,500
And of course, the methods that
we develop have to scale to these

126
00:05:53,500 --> 00:05:57,190
massive data sets that we're talking
about, different kinds of data.

127
00:05:57,190 --> 00:06:01,310
Be able to be really useful and
robust as you are analyzing the data.

128
00:06:01,310 --> 00:06:04,980
And honestly, it shouldn't require

129
00:06:04,980 --> 00:06:09,150
a PhD in machine learning to be able to
access the systems and be able to use it.

130
00:06:09,150 --> 00:06:12,140
So part of the challenge that
we have to do is be able to make

131
00:06:12,140 --> 00:06:14,220
these algorithms robust and
these methods robust.

132
00:06:14,220 --> 00:06:16,898
So that a wide range of people,
even those that don't

133
00:06:16,898 --> 00:06:20,409
have the most theoretical background
are able to build the smartest,

134
00:06:20,409 --> 00:06:22,863
coolest applications
using machine learning.

135
00:06:22,863 --> 00:06:24,305
That's exciting to me as well.

136
00:06:24,305 --> 00:06:25,209
>> Absolutely.

137
00:06:25,209 --> 00:06:29,239
[MUSIC]