1 00:00:00,900 --> 00:00:07,380 In this session, we want to see what's the effect of choosing different numbers of neurons in our head 2 00:00:07,380 --> 00:00:08,010 and linger. 3 00:00:08,880 --> 00:00:11,030 Now, let's take a look at this one. 4 00:00:11,040 --> 00:00:18,240 I'm going to use X1 and why two X1 has 50 samples and why one is equal to sign us up x1. 5 00:00:18,480 --> 00:00:20,520 This is over target to Ivon. 6 00:00:20,790 --> 00:00:22,800 We're looking for a sign of X. 7 00:00:23,730 --> 00:00:25,180 I already plotted them. 8 00:00:25,380 --> 00:00:31,710 Now let's just call a tool or simply click on neural net fitting tools. 9 00:00:32,820 --> 00:00:40,530 Here is of our neural network and as you can see, the activation function of each layer is here for 10 00:00:40,530 --> 00:00:42,230 this one forever hidden layer. 11 00:00:42,240 --> 00:00:49,260 We are using the sigmoid activation function and we use this one whenever we want to have a pattern 12 00:00:49,260 --> 00:00:50,190 recognition. 13 00:00:50,250 --> 00:00:59,430 We just want to classify them into different sets and for our altered layer, it just used linear function. 14 00:00:59,940 --> 00:01:03,510 So click on next and let's use of our data again. 15 00:01:03,540 --> 00:01:08,910 Why one four of our targets and of our inputs are X1. 16 00:01:09,150 --> 00:01:10,980 Each of them has 50 samples. 17 00:01:10,980 --> 00:01:12,360 Let's just click the next. 18 00:01:12,810 --> 00:01:17,460 And I'm not going to change the validation and testing percentage for now. 19 00:01:17,460 --> 00:01:19,740 Let's just remain midwifing. 20 00:01:20,580 --> 00:01:21,270 Here it is. 21 00:01:21,480 --> 00:01:26,580 I want to see the effect of different number of hidden neurons in our system. 22 00:01:26,820 --> 00:01:33,840 As I already explained, we can change the number of neurons in our hidden layer, but we cannot change 23 00:01:33,840 --> 00:01:36,850 the number of neurons in our output layer. 24 00:01:37,200 --> 00:01:42,830 The reason is we have only one output so we can not have more than one neuron here. 25 00:01:43,230 --> 00:01:51,660 If we had two outputs, then we could have two neurons and number of neurons for other outputs and there 26 00:01:51,660 --> 00:01:54,470 is equal to number of our outputs. 27 00:01:54,780 --> 00:01:57,840 But in this part, let's change it to one. 28 00:01:57,840 --> 00:02:03,990 Let's start with one and then see what would be the effect on our network. 29 00:02:04,530 --> 00:02:08,430 Click on next and then click on train. 30 00:02:11,200 --> 00:02:19,690 OK, the validation checks stopped our training, let Assad be checking the performance, OK is of our 31 00:02:19,690 --> 00:02:22,420 performance and this is the validation. 32 00:02:22,420 --> 00:02:25,680 We just start to failing and finally stopped our training. 33 00:02:26,170 --> 00:02:28,210 I'm going to check the feting. 34 00:02:28,540 --> 00:02:29,830 Let's take a look at this one. 35 00:02:30,880 --> 00:02:32,680 Can you guess what is this? 36 00:02:32,680 --> 00:02:40,060 By looking at this one, we can see this is a sigmoid and this is just like reverse of a sigmoid. 37 00:02:40,240 --> 00:02:45,340 And the reason that we can see it here, because one of the weights was negative. 38 00:02:45,760 --> 00:02:51,350 So it just got reversed and it just fitted to our data. 39 00:02:51,370 --> 00:02:52,710 Here are the data. 40 00:02:52,720 --> 00:02:54,100 Here are the samples. 41 00:02:54,610 --> 00:02:58,330 We were expecting something like that, like a sign of Suape lot. 42 00:02:58,510 --> 00:03:03,640 The best feting that the network was able to give us is this one. 43 00:03:03,710 --> 00:03:08,980 Maybe you ask your question why it just didn't took it up to here. 44 00:03:09,460 --> 00:03:17,200 If you start doing this sigmoid from here, then it can not satisfied samples. 45 00:03:17,200 --> 00:03:25,870 But by choosing and a middle this path, it can just satisfy all the samples in the same range to me. 46 00:03:25,870 --> 00:03:28,390 Just retrain the nets for this time. 47 00:03:28,390 --> 00:03:35,080 You're going to choose to narrowings for our hidden layer and see what will happen next. 48 00:03:36,190 --> 00:03:39,520 OK, here it's the training result. 49 00:03:40,090 --> 00:03:43,350 Again, the validation check has stopped our training. 50 00:03:43,360 --> 00:03:44,500 Let's see the fitting. 51 00:03:45,280 --> 00:03:46,780 OK, here it is. 52 00:03:46,840 --> 00:03:49,480 I remember we have only two narratives. 53 00:03:49,780 --> 00:03:52,210 As you can see, we have already two. 54 00:03:52,510 --> 00:03:57,610 This is one sigmoid still here and this is another sigmoid. 55 00:03:58,300 --> 00:04:06,850 Our network tried its best to fit into the data by using two sigmoid activation function and we can 56 00:04:06,850 --> 00:04:08,620 already see these two. 57 00:04:08,680 --> 00:04:11,140 This is one and this is another one. 58 00:04:11,410 --> 00:04:13,630 This one has a negative rate. 59 00:04:13,630 --> 00:04:14,950 That's why it's reversed. 60 00:04:14,950 --> 00:04:19,210 But this one is just look like a sigmoid and tonja and hyperbolic. 61 00:04:20,710 --> 00:04:31,060 OK, let's just increase the number of neurons to three and see how can our network behave with three 62 00:04:31,270 --> 00:04:32,890 neurones in a hidden layer. 63 00:04:33,550 --> 00:04:36,280 So let's go to the fitting tools and see the result. 64 00:04:36,700 --> 00:04:43,320 OK, the result now it's better with even tweaking around, we were able to have a sign in shape and 65 00:04:43,330 --> 00:04:45,430 identify the three neurons. 66 00:04:45,760 --> 00:04:47,620 Maybe this is one of them. 67 00:04:48,280 --> 00:04:50,110 This is one sigmoid. 68 00:04:50,620 --> 00:04:54,100 This is, I can say the second sigmoid. 69 00:04:54,280 --> 00:04:58,570 And from here to here, this is our third sigmoid. 70 00:04:58,960 --> 00:05:09,370 By changing the weights and testing it again, our network was able to feed the best that it can with 71 00:05:09,370 --> 00:05:12,910 using Htwe Sigmoid and three narrowings. 72 00:05:14,440 --> 00:05:21,040 Now let's increase the numbers to ten and see what would happen if we have pin neurons in our hidden 73 00:05:21,040 --> 00:05:21,580 layer. 74 00:05:22,750 --> 00:05:25,930 Just click on train and see the result. 75 00:05:26,200 --> 00:05:27,700 Go straight to the fitting. 76 00:05:28,510 --> 00:05:30,370 Got OK. 77 00:05:30,730 --> 00:05:32,110 Here we have this line. 78 00:05:32,110 --> 00:05:32,710 This way. 79 00:05:33,610 --> 00:05:37,300 We have some errors here and it's not a very good feed here. 80 00:05:37,300 --> 00:05:40,240 Something happened which we really don't want it. 81 00:05:40,570 --> 00:05:48,910 Sometimes giving more Naranj to other network can mislead the behavior of our network. 82 00:05:48,910 --> 00:05:55,690 It's not always good to choose a very high number of neurons, but it's very important to calculate 83 00:05:55,690 --> 00:06:01,660 the correct number of neurons in each layer and then choose it based on your data. 84 00:06:02,230 --> 00:06:08,050 So let's check what will happen if I have, let's say, 20 neurons. 85 00:06:08,920 --> 00:06:11,530 Twenty neurons in our network. 86 00:06:12,580 --> 00:06:13,510 Click on train. 87 00:06:15,310 --> 00:06:16,570 Let's see what happened. 88 00:06:18,490 --> 00:06:19,750 Oh, OK. 89 00:06:19,750 --> 00:06:20,800 Here it is. 90 00:06:21,460 --> 00:06:24,760 We just ask our network to learn the pattern. 91 00:06:24,760 --> 00:06:26,980 Using twenty sigmoid. 92 00:06:27,430 --> 00:06:30,370 I'm using twenty neurons here. 93 00:06:30,370 --> 00:06:32,890 That's why it's not where we curve. 94 00:06:32,890 --> 00:06:37,720 It's not look like a sinus wave and it's not a curved line. 95 00:06:38,290 --> 00:06:45,880 It has too many facilities and it has too many activation function. 96 00:06:45,880 --> 00:06:52,090 It just calls this shape and we can say the network has been overtrained. 97 00:06:53,820 --> 00:07:02,460 Sometimes when we give too many Klip man too many facilities to an A student, the result can be negative 98 00:07:02,730 --> 00:07:07,350 and the student might be overtrain because he has too many facilities. 99 00:07:07,350 --> 00:07:12,060 It has too many men that it can cause over training. 100 00:07:12,520 --> 00:07:16,140 So it's not a very good thing to choose to many neurons. 101 00:07:16,290 --> 00:07:21,780 But the question is, how can we know what's the best number of neurons? 102 00:07:21,780 --> 00:07:23,330 Four of our head and layer. 103 00:07:23,910 --> 00:07:26,880 We already saw that in this example. 104 00:07:26,880 --> 00:07:36,690 One, Neron is not enough to Neroni is not good, treatment is good, ten is a bit too much and 20 neurons 105 00:07:36,870 --> 00:07:38,490 actually not very good. 106 00:07:38,490 --> 00:07:39,740 And this is not acceptable. 107 00:07:39,750 --> 00:07:41,900 This not look like a sign of sweet. 108 00:07:42,420 --> 00:07:49,710 It's very important to know the best number of neurons for training and network based on our samples 109 00:07:49,710 --> 00:07:51,000 and data. 110 00:07:51,600 --> 00:07:57,870 In the next session I will explain how can we calculate the best number of neurons when we want to train 111 00:07:58,050 --> 00:07:59,490 our neural network?