WEBVTT

1
00:00:00.080 --> 00:00:03.030
- Even though symbolic reasoning, machine learning,

2
00:00:03.030 --> 00:00:06.040
and artificial neural networks are all very different,

3
00:00:06.040 --> 00:00:08.090
they're still considered part of AI.

4
00:00:08.090 --> 00:00:11.030
That means if you're starting an AI project,

5
00:00:11.030 --> 00:00:12.020
you might have to figure out

6
00:00:12.020 --> 00:00:14.090
which one of these approaches works best.

7
00:00:14.090 --> 00:00:17.010
Symbolic reasoning can perform well

8
00:00:17.010 --> 00:00:19.030
if you have an abstract problem,

9
00:00:19.030 --> 00:00:22.040
but also know the steps to find a solution.

10
00:00:22.040 --> 00:00:24.000
Machine learning is useful

11
00:00:24.000 --> 00:00:26.070
when you have to tinker to get the solution.

12
00:00:26.070 --> 00:00:28.020
You don't really know the steps

13
00:00:28.020 --> 00:00:30.050
and you have to look for larger patterns

14
00:00:30.050 --> 00:00:33.020
to come up with the likeliest answer.

15
00:00:33.020 --> 00:00:36.020
This approach requires several iterations

16
00:00:36.020 --> 00:00:38.000
to arrive at a conclusion.

17
00:00:38.000 --> 00:00:41.010
Some researchers even mix these two approaches.

18
00:00:41.010 --> 00:00:44.040
They use symbolic reasoning to come up with some constraints

19
00:00:44.040 --> 00:00:46.010
and then use machine learning

20
00:00:46.010 --> 00:00:49.010
to experiment with different answers.

21
00:00:49.010 --> 00:00:50.070
When you start your project,

22
00:00:50.070 --> 00:00:54.010
you'll have to decide which approach would work best.

23
00:00:54.010 --> 00:00:56.060
You might even try to balance the two.

24
00:00:56.060 --> 00:00:58.010
Think of it this way.

25
00:00:58.010 --> 00:01:00.060
Let's take something like a ceiling fan.

26
00:01:00.060 --> 00:01:04.030
Kids and adults alike can easily recognize one.

27
00:01:04.030 --> 00:01:06.030
Even though there are many different types,

28
00:01:06.030 --> 00:01:09.000
they usually follow a similar pattern.

29
00:01:09.000 --> 00:01:10.070
There are two or more blades

30
00:01:10.070 --> 00:01:14.050
spinning on the ceiling as a way to increase air flow.

31
00:01:14.050 --> 00:01:16.000
Now let's imagine you want to create

32
00:01:16.000 --> 00:01:18.020
an artificial intelligent program

33
00:01:18.020 --> 00:01:21.030
that could identify a ceiling fan.

34
00:01:21.030 --> 00:01:23.060
You could try to use symbolic reasoning.

35
00:01:23.060 --> 00:01:26.010
This is much closer to how most humans

36
00:01:26.010 --> 00:01:27.090
would perceive the fan.

37
00:01:27.090 --> 00:01:32.080
It relies on human experts to fully describe the fan.

38
00:01:32.080 --> 00:01:35.080
An expert would plug in some of the key attributes.

39
00:01:35.080 --> 00:01:37.010
If it's on the ceiling,

40
00:01:37.010 --> 00:01:39.070
then it's much more likely to be a ceiling fan.

41
00:01:39.070 --> 00:01:40.070
You could also plug in

42
00:01:40.070 --> 00:01:43.050
some of the names of the top manufacturers.

43
00:01:43.050 --> 00:01:45.040
Something like if it's on the ceiling

44
00:01:45.040 --> 00:01:49.010
and it has this symbol on it that says American Fans,

45
00:01:49.010 --> 00:01:52.010
then it's much more likely to be a ceiling fan.

46
00:01:52.010 --> 00:01:55.030
You could go through it this way step by step,

47
00:01:55.030 --> 00:01:58.020
hard coding most of the reasoning.

48
00:01:58.020 --> 00:02:00.000
The second option you could try

49
00:02:00.000 --> 00:02:03.050
is to use a deep learning artificial neural network.

50
00:02:03.050 --> 00:02:04.060
With this approach,

51
00:02:04.060 --> 00:02:08.010
you want to pump as much data as you can through the network.

52
00:02:08.010 --> 00:02:11.070
First, you can train the machine to recognize fans

53
00:02:11.070 --> 00:02:14.050
through images or videos of fans.

54
00:02:14.050 --> 00:02:16.050
Then you could attach a sensor

55
00:02:16.050 --> 00:02:18.070
to detect changes in wind speed.

56
00:02:18.070 --> 00:02:20.050
That way, the machine could detect

57
00:02:20.050 --> 00:02:23.010
the patterns of air moving through the room.

58
00:02:23.010 --> 00:02:24.060
You could also feed in

59
00:02:24.060 --> 00:02:27.040
terabytes of audio data into the network

60
00:02:27.040 --> 00:02:31.000
so the machine could recognize the whooshing of moving air

61
00:02:31.000 --> 00:02:33.020
or the hum of the motor.

62
00:02:33.020 --> 00:02:36.080
You'll have your neural network process all this data.

63
00:02:36.080 --> 00:02:38.080
Then it will create a model you can use

64
00:02:38.080 --> 00:02:40.040
to make an accurate guess

65
00:02:40.040 --> 00:02:44.000
about whether or not you're looking at a ceiling fan.

66
00:02:44.000 --> 00:02:45.040
Each of these approaches

67
00:02:45.040 --> 00:02:48.090
has its own advantages and disadvantages.

68
00:02:48.090 --> 00:02:52.060
Symbolic reasoning might work for most ceiling fans.

69
00:02:52.060 --> 00:02:54.060
Chances are if it's on the ceiling

70
00:02:54.060 --> 00:02:56.010
and it has two or more blades,

71
00:02:56.010 --> 00:02:57.080
then it's a ceiling fan.

72
00:02:57.080 --> 00:03:01.020
But if it's misidentified or strangely designed,

73
00:03:01.020 --> 00:03:04.010
then you won't get any type of answer.

74
00:03:04.010 --> 00:03:06.080
Machine learning also has its advantages.

75
00:03:06.080 --> 00:03:09.060
It could zero in on the perfect pattern.

76
00:03:09.060 --> 00:03:12.040
It will process a mixture of images,

77
00:03:12.040 --> 00:03:14.060
sounds, and wind movements.

78
00:03:14.060 --> 00:03:16.060
Once you fed it enough data,

79
00:03:16.060 --> 00:03:18.040
the artificial neural network

80
00:03:18.040 --> 00:03:22.010
would probably be able to identify ceiling fans

81
00:03:22.010 --> 00:03:24.030
better than most humans.

82
00:03:24.030 --> 00:03:26.080
It also has the advantage of not relying

83
00:03:26.080 --> 00:03:29.000
on the human-friendly data.

84
00:03:29.000 --> 00:03:31.070
If it's using sounds and wind motions,

85
00:03:31.070 --> 00:03:33.070
then it will have extra abilities

86
00:03:33.070 --> 00:03:36.090
beyond what humans can perceive.

87
00:03:36.090 --> 00:03:40.040
For one, you'll be able to identify ceiling fans

88
00:03:40.040 --> 00:03:45.010
24 hours a day and even in complete darkness.

89
00:03:45.010 --> 00:03:48.000
You should also keep in mind that symbolic reasoning

90
00:03:48.000 --> 00:03:49.030
would be the only approach

91
00:03:49.030 --> 00:03:52.000
if you don't have massive amounts of data.

92
00:03:52.000 --> 00:03:53.090
It's the only approach you can take

93
00:03:53.090 --> 00:03:57.070
when you don't have images, sounds, and wind speeds.

94
00:03:57.070 --> 00:04:00.080
Even for this relatively simple AI program,

95
00:04:00.080 --> 00:04:02.050
you can see how you might have

96
00:04:02.050 --> 00:04:05.040
two entirely different challenges.

97
00:04:05.040 --> 00:04:08.070
You might have one program using symbolic reasoning

98
00:04:08.070 --> 00:04:10.070
that took a long time to set up,

99
00:04:10.070 --> 00:04:13.030
but doesn't rely on outside data

100
00:04:13.030 --> 00:04:15.010
or you could have another program

101
00:04:15.010 --> 00:04:17.040
that relies on terabytes of data

102
00:04:17.040 --> 00:04:20.090
and may need to be tweaked and updated.

103
00:04:20.090 --> 00:04:22.080
When you start an AI program,

104
00:04:22.080 --> 00:04:24.040
it's important to understand

105
00:04:24.040 --> 00:04:27.030
the trade-offs of each of these approaches.

106
00:04:27.030 --> 00:04:29.040
Then you could think about the best way

107
00:04:29.040 --> 00:04:33.000
to help your machine think about the challenge.