1 00:00:00,670 --> 00:00:04,360 OK, now we proceed for a lesson to still project one. 2 00:00:04,840 --> 00:00:09,940 For this, we will be discussing on recognizing gestures with light sensors. 3 00:00:10,120 --> 00:00:14,620 But for now, we'll be tackling on the theory and data collection. 4 00:00:15,910 --> 00:00:21,580 If you actually think about it, the working principle of this project is very simple. 5 00:00:22,330 --> 00:00:27,370 The different gestures about the light sensor will block different amounts of light for different lengths 6 00:00:27,370 --> 00:00:27,880 of time. 7 00:00:29,140 --> 00:00:30,940 High volumes will be shown first. 8 00:00:31,780 --> 00:00:36,650 Nothing above the sensor, then will show low values for the time. 9 00:00:36,670 --> 00:00:44,710 Rock is above the sensor, then high values again each time a finger in paper passes above the sensor. 10 00:00:45,520 --> 00:00:50,140 We will get a value that these high, low, high, low, high, low, high, low. 11 00:00:51,160 --> 00:00:57,940 So in this lesson, we are going to train and deploy a simple neural network for classifying Rock-Paper-Scissors 12 00:00:57,940 --> 00:01:00,700 gestures with just a single light sensor. 13 00:01:01,570 --> 00:01:08,170 There is actually a high variance in speed and amplitude of values from sensor, which makes it a great 14 00:01:08,170 --> 00:01:13,870 case for using machine learning model and not hand-crafted algorithm for the task. 15 00:01:14,920 --> 00:01:21,580 So to prepare, we need to make sure that you have Arduino IDE installed and working at Real Terminal. 16 00:01:22,300 --> 00:01:23,590 Then if that's accurate? 17 00:01:23,770 --> 00:01:29,470 Create an account on each impulse website and make a new project there on this path. 18 00:01:29,980 --> 00:01:35,380 You can forward any type of sensor data to each impulse platform with data for work. 19 00:01:35,950 --> 00:01:42,940 Now here you can see here it is actually pseudo code for forwarding arbitrary sensor data from device. 20 00:01:43,930 --> 00:01:50,140 On this illustration, on the left and on the right, then if you have multiple sensor values in one 21 00:01:50,140 --> 00:01:56,740 pocket, each sensor says here values should be separated with the comma or double character. 22 00:01:56,800 --> 00:02:05,110 And then at the end of the packet, it is denoted by a new line character so you can just use serial 23 00:02:05,650 --> 00:02:12,760 dot, bring in for less value in a packet or send the new line characters separately, as seen above 24 00:02:12,970 --> 00:02:16,810 example, serial point print, for example. 25 00:02:17,500 --> 00:02:23,170 On the right side of this illustration, you could see this is actually. 26 00:02:24,340 --> 00:02:28,130 Data from a three axis accelerometer. 27 00:02:28,900 --> 00:02:35,020 And in this project, we just have one sensor that sends one value in each pocket. 28 00:02:35,620 --> 00:02:46,570 So on this slide, we need to set sample to ten 10000 milliseconds or ten seconds and then suppose for 29 00:02:46,570 --> 00:02:51,220 each gesture, such as waving the hand in the vicinity of the wheel terminal. 30 00:02:52,000 --> 00:02:59,860 Then once you uploaded the go to your terminal, make sure to run edge impulse data for wider in command 31 00:02:59,860 --> 00:03:02,830 prompt and log in with your edge impulse credentials. 32 00:03:03,160 --> 00:03:07,120 And after that, you are ready to receive data in an impulse dashboard. 33 00:03:07,630 --> 00:03:13,570 Then go to Data Acquisition tab and you should see your device there, and then you will be able to 34 00:03:13,570 --> 00:03:17,470 see the data that was collected after the samples have been taken. 35 00:03:18,760 --> 00:03:23,150 You need to make sure that the data is correct before you start collecting the next sample, actually. 36 00:03:24,040 --> 00:03:25,720 OK, now on this slide. 37 00:03:26,380 --> 00:03:32,560 This is actually a small dataset, but we also have a tiny neural network, so under fitting it's more 38 00:03:32,560 --> 00:03:33,820 likely than I'll refer to. 39 00:03:33,820 --> 00:03:42,490 In this particular case, underfunding means that a statistical model or a machine learning algorithm 40 00:03:42,490 --> 00:03:48,880 is said to be under 15 when you can configure or figuring out the underlying trend in this data. 41 00:03:49,150 --> 00:03:56,560 Now this can happen when the model size is somehow too small to come up with a general rule for data 42 00:03:56,560 --> 00:04:01,660 that has a lot of variety and noise and then four feet over fitting. 43 00:04:02,650 --> 00:04:10,060 This means that a statistical model is said to be over fitted when it starts to learn from the noise 44 00:04:10,060 --> 00:04:11,410 and mistakes in our dataset. 45 00:04:11,980 --> 00:04:13,730 This is called overfitting. 46 00:04:14,470 --> 00:04:19,570 You can do this when you have a big model and a small amount of data. 47 00:04:20,140 --> 00:04:26,830 The model can learn by heart all the data points without having to think about how they all relate to 48 00:04:26,830 --> 00:04:27,340 each other. 49 00:04:28,000 --> 00:04:34,690 So people who collect samples should make sure the model can generalize better by having a lot of different 50 00:04:34,690 --> 00:04:35,590 things to look at. 51 00:04:36,160 --> 00:04:41,710 Now, this could include things like samples that are moving in different directions or speeding away 52 00:04:41,710 --> 00:04:42,490 from the sensor. 53 00:04:43,090 --> 00:04:48,970 In general, the network only learns from the data that is in the dataset. 54 00:04:49,690 --> 00:04:56,320 If the only data you have is gestures moving from left to right above the sensor, you shouldn't expect 55 00:04:56,320 --> 00:05:02,620 the network to be able to recognize gestures moving from like the left or up and down when it's been 56 00:05:02,620 --> 00:05:07,470 trained already for left or right and on expansion. 57 00:05:08,230 --> 00:05:11,170 We need to collect more data samples in different light could.