1 00:00:11,800 --> 00:00:16,900 In this lecture we are going to summarize everything we learned in this section of the course. 2 00:00:16,900 --> 00:00:21,330 This section was all about facial recognition with Siamese networks. 3 00:00:21,370 --> 00:00:26,620 One thing you saw in this section was that while the theory was relatively simple the implementation 4 00:00:26,620 --> 00:00:33,640 was complex not because of the model itself but because of all the data processing Siamese that works 5 00:00:33,640 --> 00:00:35,750 by themselves are quite simple. 6 00:00:35,830 --> 00:00:41,710 We passed two images through the same CNN and this gives us a face embedding for both images. 7 00:00:41,710 --> 00:00:46,630 Then we would like to know the distance between these two embedding is just like we do with word embedding 8 00:00:46,630 --> 00:00:54,640 meetings that tell us how similar the two faces are if the two faces are so similar that the distance 9 00:00:54,640 --> 00:01:00,220 between the embedding is less than some threshold and we predict that they are the same person. 10 00:01:00,220 --> 00:01:04,750 If the faces are from different people the distance between the embedding should be greater than the 11 00:01:04,750 --> 00:01:09,680 threshold so that we can predict that they are different people in reality. 12 00:01:09,700 --> 00:01:15,490 Suppose you are running some kind of facial recognition database at your company instead of always passing 13 00:01:15,490 --> 00:01:22,960 in two faces at a time something simpler you could do is just store the embedding themselves in a database. 14 00:01:23,020 --> 00:01:28,750 So for all the employees of the company or all of the students in a school don't store their images 15 00:01:28,810 --> 00:01:35,050 just store there and then things then when it's prediction time you only have to calculate the embedding 16 00:01:35,350 --> 00:01:41,140 for the incoming image and compare that to all the existing embedding in your database to determine 17 00:01:41,200 --> 00:01:47,610 which person it is. 18 00:01:47,760 --> 00:01:52,590 Next we looked at all the data pre processing steps of which there were many. 19 00:01:52,590 --> 00:01:57,420 First we loaded in the data and split the images up into train and test. 20 00:01:57,420 --> 00:02:03,210 It was important to realize that the data being fed into the neural network itself not be split up but 21 00:02:03,210 --> 00:02:06,960 rather we should split up the images to start with. 22 00:02:06,960 --> 00:02:13,050 This is so that none of the images in the train set will appear in the test set and this is because 23 00:02:13,050 --> 00:02:18,930 the actual data set we train the neuron that recon simply consists of the same images being paired up 24 00:02:18,930 --> 00:02:21,950 amongst themselves multiple times. 25 00:02:22,020 --> 00:02:28,260 We don't start with a lot of images but we end up with a lot of training samples which grow quite dramatically 26 00:02:28,540 --> 00:02:31,980 with the number of images we have. 27 00:02:32,180 --> 00:02:35,900 Next we looked at how to convert the data into pairs. 28 00:02:35,930 --> 00:02:41,120 There are many ways of doing this and sometimes it can be difficult to try and understand what someone 29 00:02:41,120 --> 00:02:42,770 else's approach is. 30 00:02:42,770 --> 00:02:46,940 So as always the most important thing is to coat it up yourself. 31 00:02:46,940 --> 00:02:51,640 This always leads to the best understanding next. 32 00:02:51,660 --> 00:02:56,760 We wrote some data generator functions so that we could iterate over these in our training loop to generate 33 00:02:56,760 --> 00:03:02,640 data on the fly and not have to store them in gigantic Redundant Arrays. 34 00:03:02,640 --> 00:03:06,840 Finally we wrote a function to help us evaluate the Siamese networks. 35 00:03:06,840 --> 00:03:13,530 In particular it would draw a histogram of the match distances versus the non match distances and prints 36 00:03:13,530 --> 00:03:16,980 out the sensitivity and specificity. 37 00:03:17,010 --> 00:03:23,240 This is better than accuracy because we have imbalanced classes due to the nature of the data. 38 00:03:23,250 --> 00:03:27,420 We will always have more samples from the negative class than the positive class.