1 00:00:01,060 --> 00:00:09,710 In this lecture we will understand the concept of a fully good we have been saying the law that a cell 2 00:00:09,920 --> 00:00:17,050 in the convolutions earlier gets information from a set of pixels or a set of cells in the previous 3 00:00:17,050 --> 00:00:25,310 previously for example that red cell in the convolution earlier is getting information from these nine 4 00:00:25,580 --> 00:00:26,060 cells. 5 00:00:26,090 --> 00:00:32,160 And this red rectangle but what does this mean. 6 00:00:32,190 --> 00:00:36,810 How is it getting the information from all these pixels. 7 00:00:36,960 --> 00:00:46,010 We have 25 pixels here and our cell here can have only one value which should be the representative 8 00:00:46,010 --> 00:00:58,540 value for these 25 pixels so we need to find a way to convert these 25 values of pixels into one value. 9 00:00:58,550 --> 00:01:07,320 This is done by using a feeder filter is a matrix of same dimensions as our window of deceptive feet. 10 00:01:08,730 --> 00:01:18,430 So if the window is five cross five figure that is also a dimension 5.5 if it is of three cross three 11 00:01:18,590 --> 00:01:24,850 feet there will also be of three cross three dimensions. 12 00:01:24,880 --> 00:01:35,190 Now we have a window of five into five pixels containing pixel value and we have a five in 2 five matrix 13 00:01:35,520 --> 00:01:46,390 containing some values we multiply each pixel value with the corresponding filter value and add all 14 00:01:46,390 --> 00:01:58,220 of these products up so the pixel value here will be multiplied with zero point for the next pixel value 15 00:01:58,340 --> 00:02:07,550 will be multiplied with zero point three and so on and all these products will be added up. 16 00:02:07,610 --> 00:02:18,000 This will give us one number and this number will represent information in these 25 pixels. 17 00:02:18,190 --> 00:02:25,210 Now the question comes how do we decide the values and dispirited. 18 00:02:25,290 --> 00:02:28,290 The answer to this is very pleasing. 19 00:02:28,290 --> 00:02:36,550 We do not have to decide these values our network will learn these values also so when we are training 20 00:02:36,550 --> 00:02:39,610 our model these values will be self learned 21 00:02:43,330 --> 00:02:52,770 not to demonstrate how it does work and how they are able to extract certain features out I have taken 22 00:02:52,980 --> 00:03:01,670 a five into five input image with zero one type pixel values and a three by three filtered 23 00:03:05,690 --> 00:03:15,740 look at this way that this data looks like a cross that is the diagonal values are one and the other 24 00:03:16,030 --> 00:03:24,280 are zero if we use this filter with a straight up one we get this output 25 00:03:27,560 --> 00:03:36,440 the data below shows you how we get this output how the vector values are multiplied and their product 26 00:03:36,440 --> 00:03:44,250 values are added up to good the first value then the next value and then the next and so on. 27 00:03:49,340 --> 00:03:58,660 This final output which we get after applying the filter is called a feature map a feature map because 28 00:03:58,870 --> 00:04:07,630 each way to highlight some feature of the input image the images on the right are demonstrating how 29 00:04:07,630 --> 00:04:12,640 particular features are highlighted by filters. 30 00:04:12,870 --> 00:04:23,160 For example if we use a vertical feet that is the middle column of this matrix is 1 1 1 and these side 31 00:04:23,160 --> 00:04:31,150 columns are 0 0 0 this type of leader transforms the image to this image. 32 00:04:33,350 --> 00:04:44,720 Notice that vertical white lines are enhanced and the rest of the image is blurred similarly if we use 33 00:04:44,720 --> 00:04:55,830 the horizontal for the that is this middle row will be 1 1 1 and top and bottom row will consist of 34 00:04:55,830 --> 00:05:00,060 zeros if we use such horizontal filter. 35 00:05:01,160 --> 00:05:03,420 We got this image. 36 00:05:03,610 --> 00:05:11,390 You can notice that horizontal white lines are highlighted and rest is blurred. 37 00:05:11,980 --> 00:05:20,340 This is what operator does a filter is a set of values which transforms the window by doing some of 38 00:05:20,340 --> 00:05:27,110 products what we get after applying of that is called a feature map. 39 00:05:27,110 --> 00:05:30,410 Each feature map has some particular feature highlighted 40 00:05:33,320 --> 00:05:41,930 so what we will do is we will use many types of filter so that each filter creates different feature 41 00:05:41,930 --> 00:05:46,340 maps containing different features. 42 00:05:46,340 --> 00:05:54,480 This means our convolution earlier is going to be a bundle of feature maps and each feature map has 43 00:05:54,480 --> 00:06:03,690 some particular highlighted feature important thing to notice here is what happens in the next lit so 44 00:06:03,810 --> 00:06:12,100 this cell in the first feature map of convolution layer to what does the see. 45 00:06:12,280 --> 00:06:21,520 Is it only this rectangle on the first feature map of previously or this rectangle on all feature maps 46 00:06:21,640 --> 00:06:31,700 in the previously nonsense is that each cell on conditional Layer 2 will be getting information of all 47 00:06:31,700 --> 00:06:40,260 the feature maps in the previously because only then can these cells combine the different features 48 00:06:40,740 --> 00:06:43,050 to find more high level features 49 00:06:45,850 --> 00:06:55,100 and summaries again for clarity we apply a filter on the previous layer of data to extract features 50 00:06:58,100 --> 00:07:06,340 the output after applying filter is called a feature map we apply many different types of reactors to 51 00:07:06,340 --> 00:07:09,830 extract many different type of features. 52 00:07:09,920 --> 00:07:18,190 This gives us a bundle of feature maps the first bundle of feature maps is called constitutional Layer 53 00:07:18,230 --> 00:07:18,490 54 00:07:21,860 --> 00:07:30,080 congressional Layer 2 works on these extracted features to extract even higher level of features 55 00:07:33,670 --> 00:07:41,670 next we are going to discuss about the input layer input layer also has multiple layers of information 56 00:07:42,790 --> 00:07:47,280 these layers are called channels we talk about channels in the next video.