1
00:00:03,070 --> 00:00:03,520
Good day.

2
00:00:03,940 --> 00:00:06,250
This is from the educational engineering tech.

3
00:00:07,030 --> 00:00:10,990
Today we'll be discussing on daily email with the new terminal.

4
00:00:12,210 --> 00:00:17,230
Forward to this presentation outline on Tiny Hamlet New Terminal.

5
00:00:18,040 --> 00:00:20,370
This is actually comprised with convalescence.

6
00:00:20,770 --> 00:00:25,960
And on this lessons we will be tackling on five projects.

7
00:00:26,470 --> 00:00:34,990
First, on the lesson, one will be giving you a short introduction on what four lesson one on introduction

8
00:00:34,990 --> 00:00:42,490
to tiny email with real terminal will be giving you a short background and discussion for what we're

9
00:00:42,490 --> 00:00:44,800
going to go through throughout this presentation.

10
00:00:45,580 --> 00:00:47,080
So what is machine learning?

11
00:00:47,950 --> 00:00:54,520
It says here that it is a branch of artificial intelligence focused on building applications that learn

12
00:00:54,520 --> 00:00:59,350
from data and improve their accuracy over time without being programmed to do so.

13
00:00:59,890 --> 00:01:06,430
The foundation of machine learning is that rather than have to be thought to do everything step by step

14
00:01:06,430 --> 00:01:13,210
machines, if they can be programmed to think like us, can learn to work by observing, classifying

15
00:01:13,210 --> 00:01:19,420
and learning from its mistakes, just like we do at deep learning is a subset of machine learning that

16
00:01:19,420 --> 00:01:24,700
utilizes artificial neural networks for learning from large amounts of data.

17
00:01:25,240 --> 00:01:31,360
So what it means is that an artificial neural network is an attempt to stimulate the network of neurons

18
00:01:31,750 --> 00:01:34,090
that make up a human brain.

19
00:01:34,810 --> 00:01:42,580
Now, the ancient art created by programming computers to behave as though they are interconnected brain

20
00:01:42,580 --> 00:01:46,420
cells in order for Ian and still learning.

21
00:01:46,990 --> 00:01:53,350
They have to have a tremendous amount of information thrown at them called a train set.

22
00:01:53,890 --> 00:01:59,350
When you're trying to teach an alien and how to differentiate, it cut from the dog.

23
00:01:59,890 --> 00:02:05,770
The training set will provide thousands of images tagged as a dog so that the network would begin to

24
00:02:05,770 --> 00:02:06,040
learn.

25
00:02:06,870 --> 00:02:13,990
Once it has been trained with significant amount of data, it will try to classify future database on

26
00:02:14,380 --> 00:02:21,520
what it thinks it's seeing or hearing, depending on the dataset throughout the different units.

27
00:02:21,940 --> 00:02:27,100
Now, during the training period, the machine's output is compared to the human provided description

28
00:02:27,460 --> 00:02:30,280
of what should be observed if there is the scene.

29
00:02:30,640 --> 00:02:33,430
The machine is validated if it's incorrect.

30
00:02:33,820 --> 00:02:39,730
It uses back propagation to adjust its learning, going back through the layers to tweak the mathematical

31
00:02:39,730 --> 00:02:42,460
equation known as deep learning.

32
00:02:42,820 --> 00:02:46,690
Now this is what it makes as a network intelligence.

33
00:02:47,230 --> 00:02:53,680
Normally, deep neural networks require rather powerful computing resources to be trained and deployed.

34
00:02:54,100 --> 00:03:01,720
However, recently a batch of Mel on the edge or embedded machine learning called Dynamic have appeared.

35
00:03:02,290 --> 00:03:08,560
It represents a technique or field of study in machine learning and embedded systems that explores which

36
00:03:08,950 --> 00:03:15,790
machine learning applications once reduced, optimized and integrated, can be run on devices as small

37
00:03:15,790 --> 00:03:17,410
as microcontrollers.

38
00:03:18,160 --> 00:03:21,580
Now what is the relevance of titanium and its importance?

39
00:03:22,300 --> 00:03:28,090
It says here that the tiny arm in the arm, as you might have guessed it, stands for machine learning.

40
00:03:28,480 --> 00:03:33,160
And in most cases nowadays it refers to deep learning.

41
00:03:33,820 --> 00:03:41,980
Tiny in dynamical means that the models are optimized to run on very low power and small footprint devices

42
00:03:42,280 --> 00:03:44,080
such as various MSU's.

43
00:03:44,710 --> 00:03:51,430
Then, this year that embedded devices come in all sorts of shapes and sizes, starting from embedded

44
00:03:51,430 --> 00:04:00,280
supercomputers such as Nvidia Jetson Exchanger SGX to the tiniest of microcontrollers, for example

45
00:04:00,700 --> 00:04:04,720
the ESB 32 or Cortex M0.

46
00:04:05,230 --> 00:04:05,710
Now why?

47
00:04:05,710 --> 00:04:11,950
Embedded on microcontrollers is put in a special category and even given its own cool name.

48
00:04:12,700 --> 00:04:18,580
Now this is because it comes with its own set of advantages and limitations.

49
00:04:19,000 --> 00:04:26,290
The attractions of titanium is in fact the misuse use are ubiquitous small consume, small amounts of

50
00:04:26,290 --> 00:04:29,080
energy and comparatively cheap.

51
00:04:29,830 --> 00:04:34,210
They are M Cortex, M0 Plus and the little seed.

52
00:04:34,210 --> 00:04:37,480
We know shuffleboard, which is built around it.

53
00:04:38,020 --> 00:04:40,840
The word easy, small as thumb.

54
00:04:41,290 --> 00:04:50,290
It consumes only one point thirty three mph of power, which means it can work for more than a hundred

55
00:04:50,290 --> 00:04:58,960
and twelve hours on a 150 m battery, much more if put in deep sleep, and it only costs as little as

56
00:04:59,290 --> 00:05:00,940
4.3 US dollars.

57
00:05:01,840 --> 00:05:08,170
Now, improved model optimizations and frameworks for running machine learning model inference on microcontrollers

58
00:05:08,680 --> 00:05:11,440
have actually made it possible to add more intelligent.

59
00:05:11,990 --> 00:05:20,350
These small devices making them more useful and smarter neural networks can now be used as microcontrollers

60
00:05:20,770 --> 00:05:26,920
such as bulges, scene recognition, for example, elephant activity or a sound of breaking glass.

61
00:05:27,160 --> 00:05:34,120
We also have hotword detection routine to activate the device with a specific phrase or even for simple

62
00:05:34,120 --> 00:05:35,350
image recognition tasks.

63
00:05:35,830 --> 00:05:42,550
They can be used to give a new life to old sensors, like using an accelerometer on a machine to look

64
00:05:42,550 --> 00:05:46,270
for problems and make sure it doesn't break down before it does.

65
00:05:47,260 --> 00:05:51,190
Or they can be used to tell different kinds of liquors apart.

66
00:05:51,190 --> 00:05:55,900
Like in this demo, Tiny Animal has a lot of different ways you can use it.

67
00:05:56,920 --> 00:06:03,550
Now, you might want to ask what about its limitation now on the main limiting factor is actually the

68
00:06:03,550 --> 00:06:06,430
rum or the size of the MCU.

69
00:06:07,000 --> 00:06:16,750
No matter how you will optimize it, you wouldn't be able to fit that YOLO 999 into a tiny microcontroller.

70
00:06:17,200 --> 00:06:20,500
Same goes as for automatic speech recognition.

71
00:06:20,930 --> 00:06:24,580
Well, simple work or voice command detection is possible.

72
00:06:25,360 --> 00:06:29,320
Open domain speech recognition is out of reach of use for now.

73
00:06:30,130 --> 00:06:36,760
Now what are the preparations we need to make now in this course will mainly be using the ear and cortex

74
00:06:36,970 --> 00:06:38,120
and for gore.

75
00:06:38,140 --> 00:06:40,090
Inside we use Terminal Development Board.

76
00:06:40,630 --> 00:06:46,090
W Terminal is a perfect tool to get started with the I o d and dynamically.

77
00:06:46,240 --> 00:06:58,180
Now it is built around the HTC M d 51 P19 chip with a R M Cortex M4 f gore running to 120 hertz, which

78
00:06:58,180 --> 00:07:04,330
is very well supported by various frameworks for Ml interference on microcontrollers.

79
00:07:04,780 --> 00:07:11,320
Now it's just you that this board also has built in light sensor microphone, programmable buttons,

80
00:07:11,860 --> 00:07:20,770
2.4 inch LCD display, accelerometer to groove ports for easy connection of more than 300 various global

81
00:07:20,770 --> 00:07:24,370
ecosystem sensors in terms of software.

82
00:07:24,400 --> 00:07:32,620
We'll be using the Arduino IDE to write software that will run on devices as well as Edge Impulse and

83
00:07:32,620 --> 00:07:36,160
TensorFlow Lite for microcontrollers to train and test models.

84
00:07:36,910 --> 00:07:42,880
It's very easy to use edge impulse as a development platform for machine learning on edge devices.

85
00:07:43,630 --> 00:07:51,670
It has a web interface and toolkit for dual dynamic pipeline, from collecting data to deploying models

86
00:07:52,540 --> 00:07:54,070
needed for our next letter.

87
00:07:55,300 --> 00:08:00,700
I will be showing you how you can use your TensorFlow Lite for microcontrollers to build your own model

88
00:08:00,700 --> 00:08:02,980
training and inference pipeline.

89
00:08:03,730 --> 00:08:10,330
Now installing Arduino, I'd make sure you get the latest version from the download page.

90
00:08:10,750 --> 00:08:13,660
You can choose between installer and the zip packages.

91
00:08:14,020 --> 00:08:19,030
We suggest you use the first one that installs directly everything you need to use Arduino software,

92
00:08:19,240 --> 00:08:20,380
including the drivers.

93
00:08:20,980 --> 00:08:27,040
And then when the download finishes, make sure to proceed with installation and please allow the driver

94
00:08:27,040 --> 00:08:28,350
installation process.

95
00:08:28,360 --> 00:08:31,600
When you get a warning from the operating system.

96
00:08:32,260 --> 00:08:39,010
So before we set on our journey to explore the possibilities of tiny email, we need to set up the working

97
00:08:39,010 --> 00:08:45,640
environment, namely the Arduino IDE Edge Impulsively and TensorFlow.

98
00:08:47,110 --> 00:08:53,560
So here are the next steps in installing Arduino, I'd made sure to choose the components to install,

99
00:08:54,040 --> 00:08:55,650
as shown in the illustration.

100
00:08:56,230 --> 00:08:59,470
Then right after that, just the installation directory.

101
00:09:00,280 --> 00:09:08,290
And we suggest that to keep the default one, the next to compile and upload code for wheel terminal,

102
00:09:08,770 --> 00:09:13,150
you will need to install wheel terminal specific tools in Arduino IDF.

103
00:09:13,420 --> 00:09:17,370
And the first step is add additional boards manager.

104
00:09:17,590 --> 00:09:18,400
You are else.

105
00:09:19,210 --> 00:09:26,140
Make sure to open your original ID, click on the file, then proceed to preferences and then copy below

106
00:09:26,140 --> 00:09:35,090
the you are l to additional board manager if you are else, and then step two install the wheel terminal

107
00:09:35,090 --> 00:09:35,620
tools.

108
00:09:36,370 --> 00:09:39,850
Click on tools and then after that, go to boards.

109
00:09:40,720 --> 00:09:46,660
And then right after clicking boards, proceed two boards manager and search their wheeled terminal

110
00:09:46,660 --> 00:09:47,890
in the board's manager.

111
00:09:48,250 --> 00:09:49,360
Then if you find it.

112
00:09:49,900 --> 00:09:53,740
Press on install and wait until the installation process finishes.

113
00:09:54,490 --> 00:09:57,430
And then lastly, select your board and port.

114
00:09:58,030 --> 00:10:05,290
You'll need to select the entry in the tools and then click on board menu that corresponds to your Arduino.

115
00:10:05,650 --> 00:10:07,210
Select the real terminal.

116
00:10:08,660 --> 00:10:15,680
For installing Edge Impossible II control local devices with this Edge Impossible II that would act

117
00:10:15,680 --> 00:10:22,370
as a proxy for devices that don't have internet connection and upload and convict local files with the

118
00:10:22,380 --> 00:10:25,730
CNI, which is called the edge impossible.

119
00:10:25,730 --> 00:10:26,000
Eight.

120
00:10:26,450 --> 00:10:33,290
So basically first install node the jazz version 10 or higher on your host computer.

121
00:10:33,680 --> 00:10:39,680
Now this is applicable for Windows users installed the additional Node.js tools when prompted.

122
00:10:39,770 --> 00:10:45,260
You may skip this setup if you already have Visual Studio 2015 or more.

123
00:10:45,770 --> 00:10:56,090
Then after that, install these Yeelight tools via the npm install edge impulse like a words you should

124
00:10:56,090 --> 00:10:58,070
have the tools available in your path.

125
00:10:58,520 --> 00:11:03,080
This is just optional in installing TensorFlow for this part.

126
00:11:03,170 --> 00:11:08,390
We will be using Python and TensorFlow for more in-depth lessons later in the course.

127
00:11:08,960 --> 00:11:15,530
Then look, the installation of TensorFlow is, like I said, it is optional, though if you can get

128
00:11:15,740 --> 00:11:21,440
access to Google collab an only environment where you can execute Python code and trainer modules.

129
00:11:22,370 --> 00:11:29,780
So to install TensorFlow on Windows first install Mini Conda, which is the Python, the Virtual Environment

130
00:11:29,780 --> 00:11:36,040
Manager, which you can download from the official website and then create a new virtual environment.

131
00:11:36,050 --> 00:11:44,210
With this code conduct eight slash and Python equals 3.8.

132
00:11:45,020 --> 00:11:49,340
Then third, install TensorFlow in deep in the environment.

133
00:11:50,510 --> 00:11:51,680
Install TensorFlow.

134
00:11:52,220 --> 00:11:58,250
Then lastly, next time you need to use TensorFlow, simply open Winik on the prompt and activate the

135
00:11:58,250 --> 00:12:01,990
virtual environment you created with the code to activate Malcolm.

136
00:12:02,750 --> 00:12:08,900
So make sure to explore ready to use demos of Edge Impulse and TensorFlow by uploading sample codes.

137
00:12:09,500 --> 00:12:12,920
First is that magic wand with TensorFlow.

138
00:12:13,280 --> 00:12:16,910
Second is that voice command detection with E-ELT.