1 00:00:00,810 --> 00:00:01,240 Hello. 2 00:00:01,260 --> 00:00:02,200 Welcome back. 3 00:00:02,220 --> 00:00:05,820 So let's see how to randomly initialize our weights. 4 00:00:05,910 --> 00:00:14,320 So I'm gonna make a copy of the last project and then I should mention that over here we gonna use this 5 00:00:14,320 --> 00:00:24,470 and uh we're gonna use a sort of random number generator but ordinarily or I should say wait till I'm 6 00:00:24,480 --> 00:00:26,760 type this first. 7 00:00:26,790 --> 00:00:32,850 Ideally we should use the random number generator of our microcontroller or so a microcontroller comes 8 00:00:32,850 --> 00:00:40,740 with a hardware random number generator and that is much more precise compared to using the C standard 9 00:00:40,740 --> 00:00:43,200 library round function. 10 00:00:43,350 --> 00:00:49,040 So later on we want to see how to use the hardware random number generator on the microcontroller. 11 00:00:49,140 --> 00:00:55,020 But for this let's just try these thunder to see um run function. 12 00:00:56,280 --> 00:01:00,240 So we said Um let's see let's see this of architecture. 13 00:01:00,240 --> 00:01:05,650 So we piece these sign ups over here sign up so we send up someone. 14 00:01:06,280 --> 00:01:08,430 Basically this is a way to matrix. 15 00:01:08,430 --> 00:01:18,060 This is a matrix a matrix of weights containing the weights and this input here interact with the weights 16 00:01:18,240 --> 00:01:24,630 within this and up 0 to produce the output for this layer and then we'll take this and its Iraq to which 17 00:01:24,690 --> 00:01:31,890 this other weights to produce this sort of output and to start off in machine then in the weights are 18 00:01:31,890 --> 00:01:38,340 often initialized to random values rather than 0 we initialize the weights to random values so we would 19 00:01:38,340 --> 00:01:44,640 have to find a way to always have a random number generator to initialize it for us sort of to keep 20 00:01:44,640 --> 00:01:51,660 the ball in the air like a sports like volley someone has to throw the ball in the air and then a player 21 00:01:51,660 --> 00:01:54,060 that gets it first you stop playing with it. 22 00:01:54,060 --> 00:02:00,390 So the initialization is like getting the Portlandia so you start somewhere randomly and then the learning 23 00:02:00,390 --> 00:02:04,880 begins from the wrong random point right. 24 00:02:05,480 --> 00:02:06,060 Um yeah. 25 00:02:06,180 --> 00:02:10,060 So I'm gonna go to the simple neural network don't see. 26 00:02:11,600 --> 00:02:20,670 And I'm gonna come over here which show creates a function called the um the way to random initialization. 27 00:02:20,670 --> 00:02:25,790 And like I said this will be improved later on when we use the hardware. 28 00:02:26,280 --> 00:02:34,010 This over here we are focusing on understanding not efficiency in these initial sections. 29 00:02:34,050 --> 00:02:35,920 The focus is on understanding. 30 00:02:36,390 --> 00:02:37,310 Right. 31 00:02:37,380 --> 00:02:38,640 So I'm gonna say Wait. 32 00:02:39,960 --> 00:02:48,210 Um random initialization does a descriptive name which random initialization. 33 00:02:48,210 --> 00:02:53,640 And this is going to take a number of factors meant to be precise. 34 00:02:53,670 --> 00:02:55,410 It's gonna take three argument. 35 00:02:55,410 --> 00:03:04,400 The first one is going to be the hidden layer lend the input lint and the um the weight matrix. 36 00:03:04,680 --> 00:03:06,830 But this implies one is hidden layer. 37 00:03:06,870 --> 00:03:07,740 Right. 38 00:03:08,310 --> 00:03:09,470 So I suppose. 39 00:03:09,600 --> 00:03:09,960 Yeah. 40 00:03:09,990 --> 00:03:12,050 This the this the hidden layer. 41 00:03:13,500 --> 00:03:14,520 So when we get here. 42 00:03:15,300 --> 00:03:17,950 Okay let's take a step by step sorry about that. 43 00:03:18,060 --> 00:03:23,390 I'm gonna come over here and see you into 32 unescorted T. 44 00:03:24,150 --> 00:03:24,900 He didn't land 45 00:03:29,990 --> 00:03:33,690 and this means the number of notes in the hidden layer and then you in that to 46 00:03:38,030 --> 00:03:39,500 and then input lent 47 00:03:45,450 --> 00:03:55,180 and then a matrix to store the the output weight that will be generated. 48 00:03:55,180 --> 00:04:03,620 So say double weight matrix and the size of this number of roses. 49 00:04:03,700 --> 00:04:11,670 He didn't length number of columns as inputs land like this. 50 00:04:12,550 --> 00:04:13,160 Right. 51 00:04:13,180 --> 00:04:15,030 So open and close over here 52 00:04:21,600 --> 00:04:25,230 okay. 53 00:04:25,440 --> 00:04:36,870 And when I put a global variable here I see a double T on the score around and I'm gonna seed it 54 00:04:43,540 --> 00:04:47,420 with a comment seed random number generator 55 00:04:50,980 --> 00:04:55,040 No not C S front I should always take him into 56 00:05:02,160 --> 00:05:08,230 it with one or whatever the reason we have in this red mark is because we've not included DC standard 57 00:05:08,230 --> 00:05:23,500 library so come over here include STV end this Tedy lib sorry that H and should be find a typo fixed. 58 00:05:23,570 --> 00:05:25,160 Good right. 59 00:05:25,750 --> 00:05:31,050 So once this is done we can call our function or we can complete our function. 60 00:05:31,320 --> 00:05:33,160 So I'm simply going to say 61 00:05:36,670 --> 00:05:38,530 for int i equals zero. 62 00:05:38,610 --> 00:05:46,250 I it's less than the number of rows which is the hidden Len I plus plus open and close. 63 00:05:46,770 --> 00:05:48,670 And then for int 64 00:05:52,680 --> 00:06:01,590 for int j equals you j is less than number of columns which is input Len 65 00:06:04,860 --> 00:06:10,050 J plus plus open and close and over here 66 00:06:14,050 --> 00:06:14,870 I generate 67 00:06:19,440 --> 00:06:20,520 random numbers 68 00:06:23,910 --> 00:06:26,390 between 0 and 1. 69 00:06:27,330 --> 00:06:34,910 So come over here see a different global variable equals round 70 00:06:37,980 --> 00:06:39,000 percentage 10 71 00:06:42,200 --> 00:06:48,470 then dethroned because divided by 10. 72 00:06:50,810 --> 00:06:51,380 And then 73 00:06:55,630 --> 00:07:00,550 I simply say wait wait. 74 00:07:00,560 --> 00:07:03,140 Matrix index. 75 00:07:03,180 --> 00:07:03,590 Hi 76 00:07:07,340 --> 00:07:10,690 next J. 77 00:07:11,330 --> 00:07:11,700 Right 78 00:07:15,430 --> 00:07:20,710 and um this doesn't have to be global. 79 00:07:20,710 --> 00:07:21,490 Bring it in here. 80 00:07:38,090 --> 00:07:38,610 Sure. 81 00:07:38,630 --> 00:07:40,160 Weather across the front. 82 00:07:41,780 --> 00:07:42,200 Okay. 83 00:07:42,240 --> 00:07:43,720 So I'm gonna grab this. 84 00:07:43,720 --> 00:07:53,250 The prototype gonna put over here the head of a put a semicolon here and we can go into our main the 85 00:07:53,240 --> 00:07:55,670 c file to test it out. 86 00:07:56,510 --> 00:08:05,160 We can start off by wiping off the current content in the main function. 87 00:08:05,630 --> 00:08:09,400 So this what we have we have our data already here. 88 00:08:09,890 --> 00:08:17,710 So I'm simply going to create a number of variables. 89 00:08:17,810 --> 00:08:18,870 I'm gonna come over here. 90 00:08:19,040 --> 00:08:21,000 Let's test a simple test. 91 00:08:21,110 --> 00:08:21,860 We'll see a double 92 00:08:26,800 --> 00:08:27,880 um. 93 00:08:28,070 --> 00:08:31,290 Let's see the names we gave it to the names you gave them. 94 00:08:31,370 --> 00:08:35,090 This is S N zero s and one s and for sign ups. 95 00:08:35,530 --> 00:08:38,140 So I'm gonna say s and I'm 0. 96 00:08:38,180 --> 00:08:40,000 How about a comment first. 97 00:08:43,320 --> 00:08:48,320 So this is input layer 2 hidden layer 98 00:08:52,560 --> 00:08:57,170 which matrix after the gonna have another one. 99 00:09:00,460 --> 00:09:06,760 You didn't layer to pull layer where it. 100 00:09:06,770 --> 00:09:09,140 Matrix right. 101 00:09:09,320 --> 00:09:10,150 So let's start with this. 102 00:09:10,160 --> 00:09:15,470 We're gonna create a buffer for it. 103 00:09:15,700 --> 00:09:28,610 So this is double S N zero and number of rows because number of rows because the number of hidden nodes 104 00:09:29,820 --> 00:09:37,320 number of columns equals number of features. 105 00:09:37,570 --> 00:09:38,980 Let's see how many hidden nodes. 106 00:09:39,250 --> 00:09:40,660 So we have three nodes here. 107 00:09:40,670 --> 00:09:42,280 Does the length and one note here. 108 00:09:43,120 --> 00:09:54,340 I'm gonna come over here include a defined statement I'll say a defined number of each I.D. mean in 109 00:09:54,340 --> 00:09:57,400 hidden nodes is three 110 00:09:59,920 --> 00:10:10,850 defined number of output nodes as one like this. 111 00:10:11,000 --> 00:10:20,920 So over here the size of this matrix is number of hidden nodes and number of features which is input 112 00:10:21,370 --> 00:10:22,210 number of inputs 113 00:10:26,480 --> 00:10:28,160 then this is synopsis. 114 00:10:28,430 --> 00:10:37,270 I was in OPs one double S and one the size of this is number of output nodes 115 00:10:41,290 --> 00:10:42,940 and number of hidden nodes 116 00:10:47,920 --> 00:10:48,370 right. 117 00:10:48,640 --> 00:10:53,620 So once this is done we can go and test this out by calling our function 118 00:10:58,330 --> 00:11:09,410 come way to random initialization and then we pass this length of hidden node number of hidden nodes. 119 00:11:09,420 --> 00:11:15,530 Then number of features and then we pass this into the synopsis zero matrix 120 00:11:20,260 --> 00:11:25,420 and then we initialize the other one which random initialization 121 00:11:30,070 --> 00:11:31,420 number of output nodes 122 00:11:35,120 --> 00:11:36,260 number of hidden notes. 123 00:11:39,630 --> 00:11:42,760 And put this in Snopes one. 124 00:11:42,870 --> 00:11:45,000 No we can go ahead and print the resort 125 00:11:53,560 --> 00:11:54,440 not sure. 126 00:11:54,460 --> 00:11:55,200 Wait. 127 00:11:55,410 --> 00:12:00,160 Use unless that loop over here. 128 00:12:03,970 --> 00:12:05,810 For int I Coursera. 129 00:12:06,250 --> 00:12:07,210 It's less than 130 00:12:13,470 --> 00:12:14,190 a number of 131 00:12:17,180 --> 00:12:17,530 no. 132 00:12:17,590 --> 00:12:18,280 Hit note. 133 00:12:21,840 --> 00:12:22,260 All right. 134 00:12:22,300 --> 00:12:26,670 Plus plus open inclusion and then 135 00:12:30,600 --> 00:12:37,640 J equals zero G is less than the number of features J. 136 00:12:37,650 --> 00:12:42,870 Plus plus open and close and what we want to do is 137 00:12:46,520 --> 00:12:47,100 print f 138 00:12:52,280 --> 00:13:03,290 percentage curve then shut up sheer index by index J and then we can print some new lines here. 139 00:13:03,290 --> 00:13:03,860 Print f 140 00:13:10,800 --> 00:13:12,190 n print another one 141 00:13:19,670 --> 00:13:20,420 right. 142 00:13:20,450 --> 00:13:29,060 Once this is done we can do the same for the second set ups which is set ups one and simply going to 143 00:13:29,060 --> 00:13:37,880 copy this and change it appropriately. 144 00:13:37,880 --> 00:13:49,910 This is an ops one and this becomes output notes and then this becomes the hidden nodes 145 00:13:53,290 --> 00:13:55,560 and is not one. 146 00:13:56,500 --> 00:14:00,620 Everything else remains the same right. 147 00:14:00,670 --> 00:14:04,380 So we can do that see what we have. 148 00:14:04,390 --> 00:14:09,130 Click over here to build Click here to download onto the board. 149 00:14:09,280 --> 00:14:10,480 Let's see to write a 150 00:14:17,060 --> 00:14:17,950 reset my board. 151 00:14:20,680 --> 00:14:22,320 And this is what we have. 152 00:14:22,540 --> 00:14:30,610 So we can see the way to matrix for forcing up 0 is 3 by 3 is 3 by 2 Sorry. 153 00:14:30,610 --> 00:14:41,740 Why must see cause we have two inputs to input 3 notes we have 2 inputs coming here 3 nodes. 154 00:14:41,890 --> 00:14:57,070 So we have 3 by 2 and the way to matrix for sign ups 1 is 1 by 3 because we have one node and then 3 155 00:14:57,100 --> 00:15:01,840 inputs entering into the single node but we should talk about dimensions later. 156 00:15:01,870 --> 00:15:09,550 Now just focus on understanding later on we have a section on how to take all of this from scratch again 157 00:15:09,950 --> 00:15:16,390 and in that section with formalized the ways in which you can tell the dimensions of the various layers 158 00:15:17,230 --> 00:15:24,580 and the matrices that they work with but thus far without case study this explanation should suffice. 159 00:15:24,580 --> 00:15:30,340 If you have any questions just send me a message on this or there is for this lesson and I'll see you 160 00:15:30,340 --> 00:15:30,850 later. 161 00:15:30,850 --> 00:15:31,530 Have a nice day.