1 00:00:00,960 --> 00:00:06,900 So in this video, we will see how functional EPA can help us build complex neural network architectures. 2 00:00:09,120 --> 00:00:15,110 We have been building symbol architectures that as input goes in to bolster the layer, which goes in 3 00:00:15,110 --> 00:00:19,710 to second, how to lower the output of which goes into the final output. 4 00:00:19,990 --> 00:00:20,210 Clear. 5 00:00:21,880 --> 00:00:24,750 Now will slightly modify this architecture. 6 00:00:25,770 --> 00:00:31,520 We will take this input and merged with the output of it. 7 00:00:31,520 --> 00:00:40,170 And layer to that is will create an additional concatenation layer where the output of the layer two 8 00:00:40,590 --> 00:00:44,390 will be there and the input layer will also be added here. 9 00:00:45,570 --> 00:00:46,830 So Andy can guard clear. 10 00:00:47,270 --> 00:00:56,100 Will have 64 values from the hood and layer two because it has sixty four units of neurons and 13 values 11 00:00:56,220 --> 00:00:57,240 from the input layer. 12 00:00:57,390 --> 00:00:59,400 Because we have 13 variables. 13 00:01:00,550 --> 00:01:07,470 So in the concatenation layer, we have 77 values, which will be input into our final output layer, 14 00:01:07,590 --> 00:01:08,760 which has a single layer on. 15 00:01:10,770 --> 00:01:15,900 So this kind of architecture is called deep and wide architecture. 16 00:01:18,030 --> 00:01:21,500 And this architecture cannot be built using a sequentially big. 17 00:01:25,660 --> 00:01:28,240 This is the functionally B.A. is used. 18 00:01:29,610 --> 00:01:31,290 No, let's go back to our studio. 19 00:01:35,290 --> 00:01:39,840 Now, let's see how we can create that architecture using functional Libya. 20 00:01:41,230 --> 00:01:42,730 The first step is same. 21 00:01:43,150 --> 00:01:48,940 We have to define the input layer input layer as those 13 variables only. 22 00:01:49,480 --> 00:01:54,550 So we define the input layer with a shape of 13 variables. 23 00:01:56,380 --> 00:02:02,850 The first production layer has only to be an alert after which will come the concatenation layer. 24 00:02:03,550 --> 00:02:07,600 So in the first prediction layer, we keep the two hidden layers. 25 00:02:08,590 --> 00:02:12,760 We specified that this layer has the input from inputs. 26 00:02:12,760 --> 00:02:15,970 Splunk does the rebuild that we define for input layer. 27 00:02:17,410 --> 00:02:21,700 Then we have to put the layers after these two hidden layers. 28 00:02:22,180 --> 00:02:27,250 We need a concatenation layer in this layer. 29 00:02:27,550 --> 00:02:29,590 The output of this. 30 00:02:32,860 --> 00:02:38,170 First output layer will be concatenated with the initial input layer. 31 00:02:40,090 --> 00:02:40,960 How do we do that? 32 00:02:41,650 --> 00:02:49,600 We create this new layer, mean output in this, we concatenate these two parts. 33 00:02:50,020 --> 00:02:52,080 The first part is the prediction func. 34 00:02:53,740 --> 00:02:55,810 This includes these three layers. 35 00:02:56,980 --> 00:03:02,290 And we also add the input func, which is this input layer. 36 00:03:04,300 --> 00:03:09,610 So these two now become the input layer of this main output. 37 00:03:10,510 --> 00:03:16,090 And this goes into the last layer, which is a dense single unit output layer. 38 00:03:19,800 --> 00:03:22,980 I hope you understand the architecture that we have defined here. 39 00:03:24,060 --> 00:03:25,370 We have this input layer. 40 00:03:25,830 --> 00:03:28,480 This is we what we stored in input funk. 41 00:03:30,330 --> 00:03:35,690 Then we created predictions funk, which contains the two hidden layers. 42 00:03:36,900 --> 00:03:46,890 Then we create main output structure in which we take the output of this prediction's, funk and the 43 00:03:46,980 --> 00:03:56,130 input layer and we concatenated to create a can godlee after which we put a output layer of one single 44 00:03:56,130 --> 00:03:56,430 neuron. 45 00:03:58,980 --> 00:04:01,320 So all this is done in this part. 46 00:04:03,700 --> 00:04:08,640 You can notice here that input slier has been used twice. 47 00:04:09,480 --> 00:04:17,760 This has been made possible because in functional EPA we are having each layer as individual part of 48 00:04:17,760 --> 00:04:18,330 the structure. 49 00:04:18,840 --> 00:04:22,710 So we can use those parts many times in this structure. 50 00:04:25,230 --> 00:04:29,340 So this input funk became a part of the predictions funk variable also. 51 00:04:30,300 --> 00:04:33,500 And it became part of the main output structure also. 52 00:04:36,180 --> 00:04:37,080 So let's run this. 53 00:04:45,000 --> 00:04:52,690 And now we will define the architecture of our model using Kiraz model function in which we will claim 54 00:04:52,690 --> 00:05:02,780 that inputs are the inputs, funk and output is to be taken from the main output structure because districted 55 00:05:02,960 --> 00:05:03,830 has everything. 56 00:05:04,550 --> 00:05:11,390 It has prediction, funk also and input funk also as a concatenated layer in its structure. 57 00:05:13,180 --> 00:05:17,630 When we done this, we have Model FM radio. 58 00:05:17,630 --> 00:05:26,360 With us is a dysfunctional eBay model and it has that complex architecture stored in it. 59 00:05:27,320 --> 00:05:30,790 Now we can figure this model again. 60 00:05:31,040 --> 00:05:33,470 We will use automats properties optimizer. 61 00:05:33,830 --> 00:05:38,450 Lost is messy and the metrics to be recorded is mean absolute. 62 00:05:41,850 --> 00:05:49,290 You can see the structure of this model funk also using somebody somebody's come on, when I done this, 63 00:05:49,410 --> 00:05:55,770 you can see here that the first layer is the input layer with 13 variables. 64 00:05:57,240 --> 00:06:01,800 Then comes a dense 64 neutron layer, which is thicker than layer one. 65 00:06:03,330 --> 00:06:05,160 It is connected to the input layer. 66 00:06:06,840 --> 00:06:10,400 Then comes another dense layer, which has 64 neurons. 67 00:06:10,980 --> 00:06:15,840 And it is connected to firstly the layer after this. 68 00:06:15,960 --> 00:06:23,820 We have a concatenation layer in which we can gardino did the output of this dense layer with the 13 69 00:06:23,820 --> 00:06:25,200 variables from input layer. 70 00:06:25,830 --> 00:06:28,810 So this concatenation layer is connected to it. 71 00:06:28,830 --> 00:06:29,610 Hidden layer two. 72 00:06:30,420 --> 00:06:31,410 And the input layer. 73 00:06:35,750 --> 00:06:40,330 Then we have the last layer, which is the output layer, which has only one neuron. 74 00:06:41,540 --> 00:06:43,970 And it is connected to the concatenation. 75 00:06:44,660 --> 00:06:48,830 You know, we can bring this model using different function. 76 00:07:00,480 --> 00:07:06,720 And we can compare the best performance of our normal model with a normal beep architecture. 77 00:07:07,490 --> 00:07:12,180 What is this complex functional model with wide and deep architecture? 78 00:07:13,410 --> 00:07:15,330 So let's run these commands. 79 00:07:18,700 --> 00:07:24,740 To see that the functional model that is the complex model has a test loss of 29. 80 00:07:25,060 --> 00:07:28,640 Whereas a normal model has had a test loss of 32. 81 00:07:31,090 --> 00:07:37,240 Similarly, the mean absolute error for our functional model that is this complex model that we have 82 00:07:37,240 --> 00:07:43,810 created is forty point three, whereas the mean absolute better for the simple model that we created 83 00:07:43,930 --> 00:07:45,700 earlier was four point forty seven. 84 00:07:46,810 --> 00:07:50,950 So there is slight improvement on both of these parameters. 85 00:07:54,130 --> 00:08:01,330 Now you have seen how to use functionally B.A. to create complex neural network architecture because 86 00:08:01,330 --> 00:08:10,840 functional IPA enables us to use several layers or several structures multiple times also in that example. 87 00:08:11,170 --> 00:08:15,780 This time we saw the regression problem, the few differences that we have. 88 00:08:15,880 --> 00:08:21,700 Integration problem is, instead of using accuracy, we were using MASC. 89 00:08:24,690 --> 00:08:31,870 What they lost function and the metric we were observing, indignation problem was mean, absolute idiot 90 00:08:33,370 --> 00:08:36,490 in the architecture of the ignition problem. 91 00:08:36,760 --> 00:08:40,930 The output layer has no activation function. 92 00:08:43,570 --> 00:08:48,400 So with this, we conclude our electron functionally B.A. and regression problems.