1
00:00:00,090 --> 00:00:01,100
In this video, I'd like to

2
00:00:01,200 --> 00:00:02,840
show you a fun and historically

3
00:00:03,390 --> 00:00:05,820
important example of Neural Network Learning.

4
00:00:06,720 --> 00:00:09,300
Of using a Neural Network for autonomous driving

5
00:00:09,870 --> 00:00:12,430
that is getting a car to learn to drive itself.

6
00:00:13,810 --> 00:00:14,980
The video that I

7
00:00:15,130 --> 00:00:16,450
showed a minute, was something

8
00:00:16,820 --> 00:00:18,290
that I've gotten from Dean Pomilieu,

9
00:00:18,470 --> 00:00:20,000
who Colleague who works

10
00:00:20,260 --> 00:00:22,000
out in Carnegie Mellon University out

11
00:00:22,140 --> 00:00:23,440
on the east coast of the United States,

12
00:00:24,460 --> 00:00:25,310
and in part of the

13
00:00:25,350 --> 00:00:27,980
video you see visualizations like

14
00:00:28,230 --> 00:00:29,930
this, and what I should tell you what the visualization

15
00:00:30,080 --> 00:00:31,170
looks like before starting to

16
00:00:31,260 --> 00:00:32,830
video.
Down here

17
00:00:33,170 --> 00:00:34,860
on the lower left is the

18
00:00:35,100 --> 00:00:36,150
view seen by the car

19
00:00:36,750 --> 00:00:37,580
of what's in front of it

20
00:00:37,840 --> 00:00:38,980
and so here you know, you will kind

21
00:00:39,130 --> 00:00:40,250
of see you know, a road that's

22
00:00:40,450 --> 00:00:41,390
maybe going a bit to

23
00:00:41,470 --> 00:00:42,100
the left and going a little bit to

24
00:00:42,670 --> 00:00:45,060
the right, and up

25
00:00:45,250 --> 00:00:47,230
here on top, this

26
00:00:47,820 --> 00:00:49,820
first horizontal bar shows the

27
00:00:49,940 --> 00:00:51,510
direction selected by the

28
00:00:51,750 --> 00:00:53,110
human driver and is the

29
00:00:53,580 --> 00:00:54,630
location of this bright

30
00:00:55,180 --> 00:00:56,830
white band that shows the

31
00:00:57,100 --> 00:00:58,490
steering direction selected by the

32
00:00:58,690 --> 00:01:00,450
human driver, where, you

33
00:01:00,680 --> 00:01:01,780
know, here, far to the left

34
00:01:02,150 --> 00:01:03,280
corresponds to steering hard left;

35
00:01:03,910 --> 00:01:05,180
here corresponds to steering hard

36
00:01:05,450 --> 00:01:06,830
to the right; and so

37
00:01:06,980 --> 00:01:08,280
this location, which is a

38
00:01:08,500 --> 00:01:09,390
little bit to the left,

39
00:01:09,720 --> 00:01:10,730
a little bit left of

40
00:01:10,890 --> 00:01:12,120
center, means that the human

41
00:01:12,280 --> 00:01:13,350
driver, at this point, was

42
00:01:13,520 --> 00:01:14,600
steering slightly to the left. 
A

43
00:01:16,020 --> 00:01:17,340
nd this second part here

44
00:01:17,880 --> 00:01:18,800
corresponds to the steering

45
00:01:19,140 --> 00:01:20,720
direction selected by the

46
00:01:20,870 --> 00:01:22,020
learning algorithm; and again, the

47
00:01:22,110 --> 00:01:23,060
location of this sort

48
00:01:23,310 --> 00:01:24,790
of white band, means the

49
00:01:24,850 --> 00:01:26,560
neural network was here, selecting

50
00:01:27,040 --> 00:01:28,300
a steering direction just slightly to

51
00:01:28,380 --> 00:01:29,440
the left and in fact,

52
00:01:29,970 --> 00:01:30,980
before the neural network starts

53
00:01:31,390 --> 00:01:33,020
learning initially, you

54
00:01:33,170 --> 00:01:34,990
see that the network outputs a

55
00:01:35,170 --> 00:01:36,410
grey band, like a

56
00:01:36,580 --> 00:01:38,500
grey uniform, grey band throughout

57
00:01:38,890 --> 00:01:40,260
this region, so the uniform

58
00:01:40,740 --> 00:01:42,210
grey fuzz corresponds to the

59
00:01:42,330 --> 00:01:43,880
neural network having been randomly

60
00:01:44,450 --> 00:01:46,180
initialized, and initially having

61
00:01:46,510 --> 00:01:47,960
no idea how to

62
00:01:48,020 --> 00:01:49,650
drive the car, or initially having

63
00:01:50,050 --> 00:01:52,500
no idea what direction to steer in.

64
00:01:52,590 --> 00:01:53,640
And it's only after it's learned

65
00:01:53,860 --> 00:01:55,290
for a while that it will then start

66
00:01:55,700 --> 00:01:57,520
to output like a solid white

67
00:01:57,770 --> 00:01:58,640
band in just a small

68
00:01:58,800 --> 00:02:00,260
part of the region corresponding

69
00:02:00,700 --> 00:02:01,870
to choosing a particular steering direction.

70
00:02:02,960 --> 00:02:04,710
And that corresponds to when a neural network.

71
00:02:05,340 --> 00:02:06,890
Becomes more confident in selecting, you know,

72
00:02:08,080 --> 00:02:09,250
a and in one

73
00:02:10,220 --> 00:02:11,560
location rather than outputting

74
00:02:11,920 --> 00:02:13,110
a sort of light gray

75
00:02:13,300 --> 00:02:14,570
fuzz, but instead outputting

76
00:02:14,970 --> 00:02:17,010
a white band that's

77
00:02:17,410 --> 00:02:19,220
more constantly selecting one steering direction.

78
00:02:21,340 --> 00:02:21,880
Alban is a system

79
00:02:22,340 --> 00:02:24,850
of artificial neural networks, that learns to steer

80
00:02:25,280 --> 00:02:26,400
by watching a person drive. Alban

81
00:02:27,590 --> 00:02:29,550
is designed to control the

82
00:02:29,670 --> 00:02:31,310
tube a modified army

83
00:02:31,920 --> 00:02:32,840
Humvee who could put

84
00:02:33,020 --> 00:02:35,200
sensors, computers and actuators

85
00:02:36,160 --> 00:02:37,800
for autonomous navigation experiments.

86
00:02:41,190 --> 00:02:42,480
The initial spec in

87
00:02:42,750 --> 00:02:44,730
configuring Alban is training in

88
00:02:46,770 --> 00:02:48,160
the training the person drives

89
00:02:48,580 --> 00:02:50,640
to be a car while Alban watches.

90
00:02:55,810 --> 00:02:58,420
Once every two seconds, Alban

91
00:02:58,690 --> 00:02:59,800
digitizes a video image

92
00:03:00,320 --> 00:03:03,260
of the road ahead, and records the person's steering direction.

93
00:03:11,790 --> 00:03:13,140
This training image is reduced

94
00:03:13,560 --> 00:03:15,260
in resolution to 30 by

95
00:03:15,470 --> 00:03:16,980
32 pixels and provided

96
00:03:17,590 --> 00:03:19,100
as input to Alban's three-layer

97
00:03:21,920 --> 00:03:21,920
network.

98
00:03:22,460 --> 00:03:25,330
Using the back propagation learning algorithm; Alban

99
00:03:25,370 --> 00:03:26,590
is training to output the same

100
00:03:26,940 --> 00:03:28,450
steering direction as the

101
00:03:28,530 --> 00:03:29,970
human driver for that image

102
00:03:33,450 --> 00:03:35,970
Initially, the network's steering response is random.

103
00:03:43,930 --> 00:03:45,010
After about two minutes of

104
00:03:45,100 --> 00:03:46,760
training, the network learns

105
00:03:47,080 --> 00:03:48,730
to accurately imitate the steering

106
00:03:49,110 --> 00:03:56,430
reactions of the

107
00:03:58,670 --> 00:04:03,440
human driver. This same

108
00:04:03,710 --> 00:04:06,440
training procedure is repeated for other road types.

109
00:04:09,940 --> 00:04:11,680
After the networks have been trained the

110
00:04:11,770 --> 00:04:12,900
operator pushes the run

111
00:04:13,200 --> 00:04:14,650
switch and often begins

112
00:04:15,050 --> 00:04:20,380
driving. 12 times

113
00:04:20,870 --> 00:04:23,010
per second, Alban digitizes an

114
00:04:23,090 --> 00:04:25,130
image and feeds it to its neural networks.

115
00:04:33,210 --> 00:04:34,980
Each network, running in parallel,

116
00:04:35,930 --> 00:04:38,140
produces a steering direction and a measure of it's

117
00:04:40,050 --> 00:04:40,050
confidence in its response.

118
00:04:46,610 --> 00:04:49,480
The steering direction

119
00:04:50,140 --> 00:04:51,120
from the most confident network.

120
00:04:52,340 --> 00:04:53,650
In this case, the network trained

121
00:04:54,000 --> 00:04:56,640
for the one-lane road is used to control the vehicle.

122
00:05:04,750 --> 00:05:07,840
Suddenly, an intersection appears ahead

123
00:05:08,310 --> 00:05:09,350
of the vehicle.

124
00:05:23,090 --> 00:05:24,450
As the vehicle approaches the intersection,

125
00:05:25,680 --> 00:05:27,740
the confidence of the one-lane network decreases.

126
00:05:38,070 --> 00:05:40,030
As it crosses the intersection, and

127
00:05:40,130 --> 00:05:41,160
the two-lane road ahead comes

128
00:05:41,440 --> 00:05:44,610
into view, the confidence of the two-lane network rises.

129
00:05:51,260 --> 00:05:53,000
When it's confidence rises, the two-lane

130
00:05:53,420 --> 00:05:54,630
network is selected to steer,

131
00:05:55,050 --> 00:05:56,780
safely guiding the vehicle

132
00:05:57,380 --> 00:05:59,030
into it's lane, on the two-lane road.

133
00:06:05,400 --> 00:06:06,670
So that was autonomous

134
00:06:07,010 --> 00:06:09,790
driving using a neural network. Of course, there are more

135
00:06:10,150 --> 00:06:11,070
recently more modern attempts

136
00:06:11,910 --> 00:06:14,000
to do autonomous driving in a few properties, in

137
00:06:14,180 --> 00:06:15,730
the U.S., in Europe, and so on.

138
00:06:16,270 --> 00:06:18,040
They're giving more robust driving

139
00:06:18,400 --> 00:06:19,760
controllers than this, but I

140
00:06:20,080 --> 00:06:21,910
think it's still pretty remarkable and

141
00:06:22,040 --> 00:06:23,190
pretty amazing how a simple

142
00:06:23,690 --> 00:06:25,440
neural network trained with back-propagation

143
00:06:26,140 --> 00:06:28,990
can, you know, actually learn to drive a car somewhat well.