1
00:00:00,00 --> 00:00:06,02
(upbeat instrumental music)

2
00:00:06,02 --> 00:00:07,01
- [Instructor] I am excited

3
00:00:07,01 --> 00:00:08,09
to give you the solution

4
00:00:08,09 --> 00:00:12,01
for the Dripping Water

5
00:00:12,01 --> 00:00:15,09
Audio Classification Model challenge

6
00:00:15,09 --> 00:00:18,01
that I'd given you.

7
00:00:18,01 --> 00:00:19,09
I actually thought about it.

8
00:00:19,09 --> 00:00:22,00
I told you about dripping water,

9
00:00:22,00 --> 00:00:26,05
and I was wondering what if we used a cup of tea,

10
00:00:26,05 --> 00:00:29,07
and made it drip?

11
00:00:29,07 --> 00:00:35,01
Machine learning with an AI is so accurate

12
00:00:35,01 --> 00:00:38,05
that if you actually had dripping water

13
00:00:38,05 --> 00:00:40,01
versus dripping tea,

14
00:00:40,01 --> 00:00:41,04
it will know the difference,

15
00:00:41,04 --> 00:00:43,06
provided you give it enough sample set.

16
00:00:43,06 --> 00:00:46,02
Anyway, coming back to a dripping water problem,

17
00:00:46,02 --> 00:00:50,06
we are looking at water dripping in a pipe,

18
00:00:50,06 --> 00:00:53,07
and you are creating an HJI device,

19
00:00:53,07 --> 00:00:57,00
and it is going to have this HJI model,

20
00:00:57,00 --> 00:00:59,07
which is going to identify the sound

21
00:00:59,07 --> 00:01:01,02
of this dripping water.

22
00:01:01,02 --> 00:01:03,02
That's what we are going here for.

23
00:01:03,02 --> 00:01:06,01
I want you to remember that though

24
00:01:06,01 --> 00:01:07,09
I'm teaching you to actually

25
00:01:07,09 --> 00:01:09,09
build a training model,

26
00:01:09,09 --> 00:01:11,04
and it is going to train,

27
00:01:11,04 --> 00:01:13,04
and then you can export it.

28
00:01:13,04 --> 00:01:14,05
When you export it,

29
00:01:14,05 --> 00:01:16,02
it is an inference model,

30
00:01:16,02 --> 00:01:18,06
so what will sit on your edge device

31
00:01:18,06 --> 00:01:19,09
is an inference AI.

32
00:01:19,09 --> 00:01:21,07
So I just want you to keep that in mind

33
00:01:21,07 --> 00:01:23,05
that when you're dealing with training AI,

34
00:01:23,05 --> 00:01:25,05
that's not what the customer is seeing.

35
00:01:25,05 --> 00:01:28,00
What the customer is seeing is the inference AI.

36
00:01:28,00 --> 00:01:29,08
Now, let's look at the demo.

37
00:01:29,08 --> 00:01:31,09
The My Solution demo.

38
00:01:31,09 --> 00:01:32,07
Okay,.

39
00:01:32,07 --> 00:01:37,06
So we have background sound.

40
00:01:37,06 --> 00:01:41,06
We made this using the mic,

41
00:01:41,06 --> 00:01:44,06
and I just recorded for 20 seconds,

42
00:01:44,06 --> 00:01:47,03
so let me delete this and start again.

43
00:01:47,03 --> 00:01:52,00
And I basically said record for 20 seconds,

44
00:01:52,00 --> 00:01:55,09
kept quiet and then I said, extract sample,

45
00:01:55,09 --> 00:01:58,02
and it extracted the samples for me.

46
00:01:58,02 --> 00:02:00,08
It's actually showing 30 samples and out of 20

47
00:02:00,08 --> 00:02:02,01
because I deleted some,

48
00:02:02,01 --> 00:02:06,00
and added more samples to be extracted.

49
00:02:06,00 --> 00:02:06,08
Okay.

50
00:02:06,08 --> 00:02:10,02
So you can add any number of background noise,

51
00:02:10,02 --> 00:02:12,00
but it expects a minimum

52
00:02:12,00 --> 00:02:14,07
of 20 in teachable machine.

53
00:02:14,07 --> 00:02:16,03
Then the bell sound,

54
00:02:16,03 --> 00:02:19,02
I literally used the mic,

55
00:02:19,02 --> 00:02:24,08
and I did what we did in our demo,

56
00:02:24,08 --> 00:02:26,07
and I said, record.

57
00:02:26,07 --> 00:02:28,07
Let me start again.

58
00:02:28,07 --> 00:02:32,00
(bell dinging)

59
00:02:32,00 --> 00:02:34,03
And then I said extract,

60
00:02:34,03 --> 00:02:36,03
and I got a bunch of samples,

61
00:02:36,03 --> 00:02:39,01
and I got about 12 samples.

62
00:02:39,01 --> 00:02:40,01
Okay.

63
00:02:40,01 --> 00:02:41,06
Then I made a new class.

64
00:02:41,06 --> 00:02:43,06
This is a dripping water class.

65
00:02:43,06 --> 00:02:45,03
I just called it water.

66
00:02:45,03 --> 00:02:46,04
Again, the name doesn't matter.

67
00:02:46,04 --> 00:02:51,04
It's just a name for this group of data,

68
00:02:51,04 --> 00:02:53,08
just named as water in this case.

69
00:02:53,08 --> 00:02:54,08
Okay.

70
00:02:54,08 --> 00:02:57,07
So again, it says you need eight minimum.

71
00:02:57,07 --> 00:03:00,03
So I'm going to use the mic,

72
00:03:00,03 --> 00:03:03,06
and I'm going to record,

73
00:03:03,06 --> 00:03:05,06
and this is dripping water.

74
00:03:05,06 --> 00:03:06,05
Okay.

75
00:03:06,05 --> 00:03:11,02
So I have a cup with water in a spoon,

76
00:03:11,02 --> 00:03:19,08
and as soon as I hit record,

77
00:03:19,08 --> 00:03:21,03
I didn't like it.

78
00:03:21,03 --> 00:03:22,05
I think I started slow,

79
00:03:22,05 --> 00:03:25,05
so I'm going to get rid of it.

80
00:03:25,05 --> 00:03:26,09
See what I did here?

81
00:03:26,09 --> 00:03:28,09
I don't need to extract the sample.

82
00:03:28,09 --> 00:03:30,03
I can just ignore it.

83
00:03:30,03 --> 00:03:31,07
So I wanted to show you,

84
00:03:31,07 --> 00:03:33,03
you don't have to be perfect.

85
00:03:33,03 --> 00:03:35,03
You can try this any number of time.

86
00:03:35,03 --> 00:03:36,09
When you extract the sample,

87
00:03:36,09 --> 00:03:38,03
then it extracts the sample.

88
00:03:38,03 --> 00:03:39,03
Even at that point,

89
00:03:39,03 --> 00:03:41,05
you didn't like it, you can d delete it.

90
00:03:41,05 --> 00:03:43,02
See, it's that simple.

91
00:03:43,02 --> 00:03:44,00
Okay.

92
00:03:44,00 --> 00:03:45,05
So I'm going to do it again,

93
00:03:45,05 --> 00:03:50,03
and I'm holding a spoon of water over a cup.

94
00:03:50,03 --> 00:03:53,00
I'm going to let it drip,

95
00:03:53,00 --> 00:03:55,03
and at the same time I'm going to press

96
00:03:55,03 --> 00:04:01,02
record two seconds.

97
00:04:01,02 --> 00:04:02,02
Okay.

98
00:04:02,02 --> 00:04:04,07
Now extract the sample.

99
00:04:04,07 --> 00:04:07,04
Now I need eight of them.

100
00:04:07,04 --> 00:04:11,03
I'm going to repeat this.

101
00:04:11,03 --> 00:04:12,08
Okay, I've got 10 samples.

102
00:04:12,08 --> 00:04:15,03
You don't have to have an exact equal number.

103
00:04:15,03 --> 00:04:19,04
I have 12 in the bell and 10 in here.

104
00:04:19,04 --> 00:04:20,03
It doesn't matter.

105
00:04:20,03 --> 00:04:22,06
It has to be kind of close.

106
00:04:22,06 --> 00:04:23,04
Okay.

107
00:04:23,04 --> 00:04:25,01
So what did we do in our training?

108
00:04:25,01 --> 00:04:28,00
We collected our samples in three classes.

109
00:04:28,00 --> 00:04:29,05
That's what we've done here.

110
00:04:29,05 --> 00:04:32,06
And then I taught you about epoch is the number

111
00:04:32,06 --> 00:04:33,09
of times the training data

112
00:04:33,09 --> 00:04:35,05
is going to be scanned.

113
00:04:35,05 --> 00:04:38,08
And if you look closely, you can see

114
00:04:38,08 --> 00:04:40,05
the sound is digitized,

115
00:04:40,05 --> 00:04:43,05
and the frequency is captured here beautifully.

116
00:04:43,05 --> 00:04:46,06
Okay, so I just press click,

117
00:04:46,06 --> 00:04:49,07
and it'll train the model,

118
00:04:49,07 --> 00:04:54,04
and it says preparing training and it's done.

119
00:04:54,04 --> 00:04:55,02
Yeah.

120
00:04:55,02 --> 00:04:56,00
Okay.

121
00:04:56,00 --> 00:04:58,02
So now as I'm talking,

122
00:04:58,02 --> 00:05:00,00
it's trying to hear my voice

123
00:05:00,00 --> 00:05:02,07
because I've turned input here

124
00:05:02,07 --> 00:05:03,08
from the mic,

125
00:05:03,08 --> 00:05:05,04
and it's trying to guess

126
00:05:05,04 --> 00:05:07,00
the sound that it is hearing.

127
00:05:07,00 --> 00:05:08,04
Is that background?

128
00:05:08,04 --> 00:05:11,01
Is that the bell or is it the water?

129
00:05:11,01 --> 00:05:13,02
As I taught you during the demo,

130
00:05:13,02 --> 00:05:15,07
this AI's world is very narrow.

131
00:05:15,07 --> 00:05:18,09
It knows only three classes of sounds,

132
00:05:18,09 --> 00:05:19,09
the background noise,

133
00:05:19,09 --> 00:05:22,01
the bell sound and the water sound.

134
00:05:22,01 --> 00:05:24,01
So anything else it hears

135
00:05:24,01 --> 00:05:26,02
is going to fit under that.

136
00:05:26,02 --> 00:05:29,04
See, if I go and go add a different sound.

137
00:05:29,04 --> 00:05:31,01
(sounds tapping)

138
00:05:31,01 --> 00:05:36,02
I'm tapping my spoon to my cup,

139
00:05:36,02 --> 00:05:38,04
and it's just trying to fit in one of them.

140
00:05:38,04 --> 00:05:40,02
Okay, now for the real test.

141
00:05:40,02 --> 00:05:41,01
I'm going to be quiet,

142
00:05:41,01 --> 00:05:45,00
and see whether it recognize the background.

143
00:05:45,00 --> 00:05:45,08
Okay.

144
00:05:45,08 --> 00:05:47,06
So 99% or 97%,

145
00:05:47,06 --> 00:05:49,01
it thinks it's background.

146
00:05:49,01 --> 00:05:50,04
That's good.

147
00:05:50,04 --> 00:05:52,05
Now I'm going to test for the dripping water.

148
00:05:52,05 --> 00:05:54,05
That is our real use case, right.

149
00:05:54,05 --> 00:05:57,03
Pipe is leaking and you need to find

150
00:05:57,03 --> 00:05:59,04
the AI needs to recognize it

151
00:05:59,04 --> 00:06:02,07
so that it can notify you

152
00:06:02,07 --> 00:06:05,09
or turn on the water, turn off the water,

153
00:06:05,09 --> 00:06:08,01
do something meaningful, some automation.

154
00:06:08,01 --> 00:06:10,03
It gives you the power to do that, right.

155
00:06:10,03 --> 00:06:17,02
Here we go.

156
00:06:17,02 --> 00:06:19,01
So the number goes up.

157
00:06:19,01 --> 00:06:22,01
It could be another viscous fluid,

158
00:06:22,01 --> 00:06:25,08
which will have a different sound when it drips.

159
00:06:25,08 --> 00:06:27,09
And so if you want to capture that,

160
00:06:27,09 --> 00:06:29,01
and you want your AI model

161
00:06:29,01 --> 00:06:30,03
to capture that,

162
00:06:30,03 --> 00:06:32,00
you want to train for that,

163
00:06:32,00 --> 00:06:34,04
and you want to create the inference model,

164
00:06:34,04 --> 00:06:35,09
which you're going to automate.

165
00:06:35,09 --> 00:06:38,09
Maybe you have a full automation done

166
00:06:38,09 --> 00:06:42,08
from the HJI device or it sends you notification

167
00:06:42,08 --> 00:06:46,02
or it's integrated into your broader workflow

168
00:06:46,02 --> 00:06:47,05
into other systems.

169
00:06:47,05 --> 00:06:49,01
So there's so many possibilities,

170
00:06:49,01 --> 00:06:50,08
but what you have learned here

171
00:06:50,08 --> 00:06:52,06
is the foundation knowledge

172
00:06:52,06 --> 00:06:55,06
of how that machine learning classification

173
00:06:55,06 --> 00:06:56,04
is done,

174
00:06:56,04 --> 00:06:59,03
how it is so easy to collect data.

175
00:06:59,03 --> 00:07:01,02
So in instead of thinking of data

176
00:07:01,02 --> 00:07:02,05
as multiple data sets,

177
00:07:02,05 --> 00:07:04,01
and then download from here and there,

178
00:07:04,01 --> 00:07:05,06
and things like that,

179
00:07:05,06 --> 00:07:06,09
it's very simple,

180
00:07:06,09 --> 00:07:10,03
and you just create different class of data,

181
00:07:10,03 --> 00:07:12,07
and you train it, you export it,

182
00:07:12,07 --> 00:07:14,02
and you've got the answer.

183
00:07:14,02 --> 00:07:15,03
So let me show the export

184
00:07:15,03 --> 00:07:19,04
to finish up the challenge solution.

185
00:07:19,04 --> 00:07:23,08
All I did is press export and I said,

186
00:07:23,08 --> 00:07:26,02
upload my model,

187
00:07:26,02 --> 00:07:28,08
and it's uploading to the cloud.

188
00:07:28,08 --> 00:07:32,09
I copy that and this is where it is going

189
00:07:32,09 --> 00:07:34,06
to put the inference model.

190
00:07:34,06 --> 00:07:37,05
This inference model that you copy,

191
00:07:37,05 --> 00:07:39,05
you could also download it,

192
00:07:39,05 --> 00:07:41,00
and you can download it

193
00:07:41,00 --> 00:07:42,08
and burn it into your edge device.

194
00:07:42,08 --> 00:07:46,04
Edge AI typically sits on the Edge,

195
00:07:46,04 --> 00:07:47,07
and that is why it's call Edge AI

196
00:07:47,07 --> 00:07:48,08
in the first place,

197
00:07:48,08 --> 00:07:50,04
and then it runs inference from there.

198
00:07:50,04 --> 00:07:52,05
So it receives data like the sound

199
00:07:52,05 --> 00:07:53,07
of a dripping water

200
00:07:53,07 --> 00:07:56,02
or whatever your model is supposed to do,

201
00:07:56,02 --> 00:07:57,06
and it'll recognize

202
00:07:57,06 --> 00:07:59,09
and then it'll say, yes,

203
00:07:59,09 --> 00:08:01,01
I identified a water.

204
00:08:01,01 --> 00:08:02,05
Now what do you do with it

205
00:08:02,05 --> 00:08:04,01
is where you build the product

206
00:08:04,01 --> 00:08:05,07
of a notification integration

207
00:08:05,07 --> 00:08:06,05
into your workflow

208
00:08:06,05 --> 00:08:08,09
or whatever you action you want to take.

209
00:08:08,09 --> 00:08:10,00
Okay.