1 00:00:02,840 --> 00:00:05,790 ‫Now, let us look at another CNN architecture. 2 00:00:06,510 --> 00:00:09,680 ‫GoogleNet was the winner of 2014 challenge. 3 00:00:10,550 --> 00:00:18,620 ‫GoogleNet had one new concept and the concept was of an Inception module. Inception module 4 00:00:18,770 --> 00:00:20,000 ‫Looks something like this. 5 00:00:22,370 --> 00:00:27,200 ‫That input to the Inception module is given to four different layers. 6 00:00:27,860 --> 00:00:31,450 ‫Three of these layers are convolutional layers. 7 00:00:31,940 --> 00:00:33,500 ‫And the fourth one is a max pool layer. 8 00:00:35,750 --> 00:00:40,420 ‫If you look at these convolutional layers, these are also one by one kernel. 9 00:00:41,120 --> 00:00:44,210 ‫That is the window is of the size of a single pixel. 10 00:00:45,140 --> 00:00:50,210 ‫Usually we have been using convolutional layers with two by two or three by two window. 11 00:00:50,990 --> 00:00:56,570 ‫But in the Inception module, you can see that windows or size one by one are used. 12 00:00:57,530 --> 00:00:59,490 ‫This one as represents the stride. 13 00:00:59,660 --> 00:01:04,240 ‫So it has a strode of one, the output of these two convolutional layers. 14 00:01:04,850 --> 00:01:09,080 ‫And this max pooling layer then went into three different convolutional layers. 15 00:01:10,100 --> 00:01:15,340 ‫The output of these four was then put into a depth concatinator. 16 00:01:16,250 --> 00:01:18,500 ‫Will not be discussing that depth concatinator here. 17 00:01:19,880 --> 00:01:22,730 ‫This whole thing is called an inception module. 18 00:01:24,380 --> 00:01:32,660 ‫And the actual architecture of GoogleNet was something like this input of images right through here. 19 00:01:32,810 --> 00:01:39,500 ‫Then there was a convolutional layer, a max pool layer, a local response normal layer, two convolutional layers 20 00:01:39,500 --> 00:01:40,070 ‫and so on. 21 00:01:40,850 --> 00:01:44,090 ‫All of these are stacked inception modules. 22 00:01:44,330 --> 00:01:45,980 ‫So this is the Inception layer. 23 00:01:46,100 --> 00:01:47,420 ‫This is Inception layer. 24 00:01:48,530 --> 00:01:55,010 ‫So many of these inception layers, the output of these goes into another set of inception layers. 25 00:01:55,490 --> 00:01:58,840 ‫And then we finally have a fully connected neural network. 26 00:02:01,740 --> 00:02:10,380 ‫So if we look at it, it is a very complex network, although it had very few training parameters for 27 00:02:10,380 --> 00:02:16,140 ‫most of the working professionals or student working in the field of data science and machine learning, 28 00:02:16,590 --> 00:02:21,660 ‫creating such architectures on their machines is not possible. 29 00:02:22,650 --> 00:02:31,620 ‫So one of the good things that is coming with our Guide US library is that we are able to use these 30 00:02:31,630 --> 00:02:35,440 ‫pre trained models for our problem. 31 00:02:38,760 --> 00:02:43,350 ‫These architectures were created to solve one particular problem. 32 00:02:44,300 --> 00:02:49,360 ‫But Tenent allow us to use these architectures for other problems. 33 00:02:49,380 --> 00:02:51,660 ‫Also, let us see how.