1 00:00:00,870 --> 00:00:04,770 Below we have discussed the individual sale. 2 00:00:04,920 --> 00:00:13,610 Now we are going to start these says to create network offsets just to avoid confusion with biological 3 00:00:13,610 --> 00:00:14,690 neuron. 4 00:00:14,750 --> 00:00:19,640 I'll be calling a neuron as perception runs only through perception. 5 00:00:19,640 --> 00:00:24,310 From now on means any artificial neuron. 6 00:00:24,440 --> 00:00:32,000 Now there are two ways we can stack cells violently or sequence in. 7 00:00:32,170 --> 00:00:43,420 Let's see what happens when this tax is badly hit as a single positron with three inputs and one output. 8 00:00:43,430 --> 00:00:52,370 Now we had a deposit brought bad luck to this cell also gets the same three inputs but it has a different 9 00:00:52,460 --> 00:00:58,690 output why do we can keep on adding more cells better to these ones. 10 00:00:58,820 --> 00:01:07,040 Maybe a third or fourth or even more than that we we'll just start getting new output or in other words 11 00:01:07,280 --> 00:01:15,920 we can predict for multiple output using the same input features for example when we are doing image 12 00:01:15,980 --> 00:01:20,670 recognition and we are trying to find out a face of a person. 13 00:01:21,020 --> 00:01:29,080 He may also want to find the x and y coordinate of that phase equal to for these two variables will 14 00:01:29,080 --> 00:01:36,530 become vital NYT although image definition needs a much more complex network. 15 00:01:36,630 --> 00:01:43,380 By giving this example I wanted to make the point that neural networks are not bound to only one output 16 00:01:44,810 --> 00:01:51,470 with the same input you can get multiple output because we can do parallel stacking of the artificial 17 00:01:51,470 --> 00:01:54,740 neurons. 18 00:01:54,740 --> 00:02:00,290 Now let's see going chilled stacking in the image above. 19 00:02:00,500 --> 00:02:09,390 We have five inputs which we input to three parallel separate now the output from this set of perceptions 20 00:02:09,690 --> 00:02:16,210 is taken and Fred as input to another set of valid perceptions. 21 00:02:16,430 --> 00:02:25,220 Here I am inputting the output of these three to these four plus approx again I take the four outputs 22 00:02:25,220 --> 00:02:30,940 of these perceptions and input these into this single perceptual. 23 00:02:31,020 --> 00:02:38,050 Lastly the singleton is giving out one single output which is the variable which we want to predict. 24 00:02:41,440 --> 00:02:50,170 So this is sequential stacking in which the output of one set of parallel least tagged neurons is sequentially 25 00:02:50,200 --> 00:02:54,220 given as input to the next set of parallels to act neurons. 26 00:02:56,640 --> 00:03:00,380 Let's first understand the benefit of doing this. 27 00:03:00,530 --> 00:03:09,390 That is why did we not just input all the five inputs into a single cell and use this output to predict 28 00:03:09,460 --> 00:03:20,250 variable like how is stacking these additional sets of neuron helpful to we have this type of beta. 29 00:03:20,310 --> 00:03:28,240 There are these two input variables maybe height and weight basis which we are trying to classify. 30 00:03:28,710 --> 00:03:36,960 If the anyone in the room is a cow or a dog so cows generally like it they have more weight and more 31 00:03:36,960 --> 00:03:38,520 heat than a dog. 32 00:03:38,640 --> 00:03:45,010 And dogs generally like it that is they are represented by the red dot. 33 00:03:45,140 --> 00:03:52,700 Now when we're classifying this sort of data we can have a linear separator that is a straight line 34 00:03:53,090 --> 00:03:54,890 to separate these two classes. 35 00:03:56,720 --> 00:04:03,310 Anything on the right side will be predicted as a call and anything on the left side will be predicted 36 00:04:03,450 --> 00:04:03,960 as a dog. 37 00:04:05,040 --> 00:04:13,790 This is the capability of a single perceptual single perception can find out the best straight line 38 00:04:14,150 --> 00:04:22,920 to classify the given data so if we have this problem using a single person wrong would suffice. 39 00:04:24,340 --> 00:04:32,500 But the situation is more complex in fact in real life situations we'd never use neural networks when 40 00:04:32,500 --> 00:04:38,770 we need to classify a photo situation as simple as this the real life situations for neural networks 41 00:04:38,980 --> 00:04:49,470 is often more complex let me complicate the example a little bit what if we wanted to classify objects 42 00:04:49,830 --> 00:04:56,770 which have this distribution so anything to the left of the first name and anything to the date of the 43 00:04:56,770 --> 00:05:06,940 second line is Class 8 or is it a dot and anything in between these two lines is Class B or if a green 44 00:05:06,940 --> 00:05:16,630 dot this type of classification situation cannot be handled by a single person on a netbook such as 45 00:05:16,960 --> 00:05:19,790 the one shown on the right can easily handle it. 46 00:05:21,230 --> 00:05:25,360 For example this bus neuron will fire. 47 00:05:25,560 --> 00:05:34,430 That is give output as one ready point lays to the left of lane one and the second neuron will give 48 00:05:34,430 --> 00:05:43,560 output as one ready point lays to divide off line to and this final neuron gives output as one when 49 00:05:43,770 --> 00:05:50,040 any one of the two inputs is one you can polish the video. 50 00:05:50,370 --> 00:05:59,220 Think about it for a couple of minutes and see how this small network is handling this special classification. 51 00:05:59,230 --> 00:06:03,630 This is the power of a neural network in the network. 52 00:06:03,630 --> 00:06:11,850 We created each neuron can focus on a particular feature of the object and not on the final output. 53 00:06:12,990 --> 00:06:20,480 The final output will be predicted based is the desert of these features in this way. 54 00:06:20,530 --> 00:06:26,890 Neural networks can do really sophisticated decision making with basic machine learning techniques such 55 00:06:26,890 --> 00:06:29,890 as linear regression cannot do with good accuracy 56 00:06:32,660 --> 00:06:34,360 before we move on. 57 00:06:34,520 --> 00:06:36,500 Let's take a minute to discuss this. 58 00:06:36,500 --> 00:06:39,630 Networks nominated. 59 00:06:39,850 --> 00:06:43,000 This is a neural network now. 60 00:06:43,050 --> 00:06:53,590 Each set of parallel neurons are called Live that first is the input layer the last is the output left 61 00:06:54,260 --> 00:07:04,080 and these in-between decoding live this network had five inputs three cells inherently and one for inherently 62 00:07:04,110 --> 00:07:10,050 two and one in the output layer so for brevity. 63 00:07:10,270 --> 00:07:14,250 This network can also be called as a 5 3 4 1. 64 00:07:14,290 --> 00:07:14,800 Network 65 00:07:18,630 --> 00:07:27,550 ultimate is that the process information in this network is flowing in only the forward detection which 66 00:07:27,550 --> 00:07:34,010 is why the network is also called a feed forward network in comparison. 67 00:07:34,050 --> 00:07:41,130 If the output of one of these cells of that layer goes back as input to end of the scale of that literally 68 00:07:41,700 --> 00:07:48,030 then it is called a cyclic network but it could have neural networks also known as audit. 69 00:07:48,020 --> 00:07:55,780 In the example of cyclic network auditing are used in natural language processing and language modeling. 70 00:07:55,970 --> 00:08:00,560 For now let's come back to standard feed for word network. 71 00:08:00,620 --> 00:08:09,330 Now you can notice here that output from this said is going to force it. 72 00:08:09,360 --> 00:08:11,700 These are not for different outputs. 73 00:08:11,700 --> 00:08:21,590 It is only one output the same output is going as input in all these pulses also note that every neuron 74 00:08:21,830 --> 00:08:30,340 in each layer is connected to every other neuron in the Edison forward layer therefore this network 75 00:08:30,340 --> 00:08:31,680 is fully connected. 76 00:08:33,120 --> 00:08:41,100 If somehow some links were missing we call it possibly connected but for most practical purposes we 77 00:08:41,100 --> 00:08:42,530 use a fully connected network. 78 00:08:45,680 --> 00:08:51,680 Before I close this LICHTER I would like to tell you that within this short span of time in which we 79 00:08:51,680 --> 00:08:59,510 covered this LICHTER We have entered devoid of deep learning such artificial neural networks is what 80 00:08:59,600 --> 00:09:02,980 deep learning is made up of basically. 81 00:09:03,150 --> 00:09:09,330 Think of this like a system which loans the relationship between input and output. 82 00:09:10,850 --> 00:09:19,190 The more layers we have in the system the more deep our system is more that is capable of establishing 83 00:09:19,260 --> 00:09:28,820 a complex relationship between input and output so I hope that you understand the basics of neural networks 84 00:09:30,010 --> 00:09:37,600 in the next lecture will go deeper and see how these networks process the output and find the optimum 85 00:09:37,810 --> 00:09:45,830 values of weight and biases to get good accuracy of prediction unanimously.