Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,390 --> 00:00:06,470 Crais final step that we're going to look at noise goes into the generator generator generates dogs 2 00:00:06,540 --> 00:00:13,090 with eyes this time and we can look more three dimensional I have ears for different. 3 00:00:13,100 --> 00:00:15,680 Too bad they have too many eyes. 4 00:00:15,680 --> 00:00:22,990 But some of them but already you know it's learned from the mistakes it made previously. 5 00:00:23,160 --> 00:00:28,560 And so now we want to train the discriminator and we're going to need another batch of dog images we're 6 00:00:28,560 --> 00:00:30,000 going to put them in. 7 00:00:30,000 --> 00:00:33,360 And there you go see what outputs you've got. 8 00:00:33,360 --> 00:00:44,180 We've got some valuable top to bottom and so what we can see now is that these values are slowly converging 9 00:00:44,180 --> 00:00:51,470 so the value of getting closer the values for the dogs or the dogs are getting bigger with time and 10 00:00:51,470 --> 00:00:57,550 it's harder it is becoming harder for the discriminator to discriminate between dogs and dogs. 11 00:00:57,550 --> 00:01:03,230 I also noticed that in the Paduan retraining the general and that's good that means these images are 12 00:01:03,230 --> 00:01:05,170 getting closer to what we want to. 13 00:01:05,210 --> 00:01:07,130 More realistic dogs. 14 00:01:07,130 --> 00:01:12,590 And again we're going to take these values Kellow can calculate the errors based on what we know they 15 00:01:12,590 --> 00:01:15,410 should be they should be zeros at the top ones at the bottom. 16 00:01:15,740 --> 00:01:20,300 And that error is going to be back propagated through the network of the discriminator. 17 00:01:20,320 --> 00:01:26,350 Weights are going to be updated and then then it's time to train the generators. 18 00:01:26,360 --> 00:01:28,930 Now we're going to train the generator. 19 00:01:28,940 --> 00:01:32,800 We're going to have to put these images into the network. 20 00:01:32,990 --> 00:01:36,500 And what happens next is we get an output. 21 00:01:36,500 --> 00:01:43,700 You can see this output is lower than it was just now but also as you can see it slowly. 22 00:01:43,760 --> 00:01:46,470 It's not as low as at the very start. 23 00:01:46,560 --> 00:01:55,100 Again this is a sign it's showing us that these images are actually getting closer to realistic dog 24 00:01:55,130 --> 00:01:56,050 images. 25 00:01:56,140 --> 00:02:01,500 Again that these valves are going to be we're going to get the values collate the error and get back 26 00:02:01,510 --> 00:02:08,260 Burkitt the error through the network of the generator and the weights there. 27 00:02:08,750 --> 00:02:10,970 So there we go that's how it works. 28 00:02:10,970 --> 00:02:16,550 These are just three steps in training in reality these like hundreds and thousands of these steps and 29 00:02:16,550 --> 00:02:19,670 then multiple airports as well. 30 00:02:19,670 --> 00:02:24,410 And as you can imagine through many many many many iterations these images are going to get better and 31 00:02:24,410 --> 00:02:31,790 better and better and better through this struggle through this confrontation between the generator 32 00:02:31,850 --> 00:02:34,100 and the discriminant. 33 00:02:34,340 --> 00:02:37,740 Additional information is definitely available. 34 00:02:37,760 --> 00:02:43,400 And one of the best place to go to is the original paper by in good fellow it's called generative adversarial 35 00:02:43,490 --> 00:02:48,420 Annetts 2014 paper which you can find archive. 36 00:02:48,530 --> 00:02:48,880 Yeah. 37 00:02:48,890 --> 00:02:51,110 And he explains everything there. 38 00:02:51,500 --> 00:02:54,570 And one more thing I wanted to mention here. 39 00:02:54,620 --> 00:02:58,850 When we were talking about the generator we just said it's a neural network. 40 00:02:58,850 --> 00:03:03,850 But this type of neural network we haven't discussed it is not even discussed in the annex. 41 00:03:03,950 --> 00:03:10,300 It's called a deconvolution or neural network in that case you'll find artificial neural networks and 42 00:03:10,300 --> 00:03:14,070 can we do illusional neural networks but not deconvolution all neural networks. 43 00:03:14,240 --> 00:03:20,090 So I just wanted to make a quick note here so there is this illusional neural network what a deconvolution 44 00:03:20,090 --> 00:03:27,350 all neural network is if you flip this upside down or run it back to front and then run it this way 45 00:03:28,460 --> 00:03:30,660 instead of the normal way you run the other way. 46 00:03:30,680 --> 00:03:35,870 Basically you start with the vector of values which is our Random signal. 47 00:03:35,880 --> 00:03:38,180 Then at the end you'll get the image. 48 00:03:38,420 --> 00:03:43,940 And if you'd like to learn more about decompositional neural networks there is some additional reading 49 00:03:43,940 --> 00:03:44,700 over here. 50 00:03:44,750 --> 00:03:52,730 We're not discussing it here because it's not the main concept of for us to focus on one focus on Ganns 51 00:03:52,760 --> 00:03:59,450 But if you'd like to learn more about D-Conn. finance then this is a paper which talks about them is 52 00:03:59,450 --> 00:04:06,260 called adaptive deconvolution networks for mid and high level feature learning by Mathieu's and others. 53 00:04:06,500 --> 00:04:14,780 So yeah that's a go to people are forged continents hopefully you enjoyed this tutorial and you now 54 00:04:14,780 --> 00:04:20,840 know how Ganns work in the background and then report seeing you back here next time. 55 00:04:20,850 --> 00:04:22,710 And until then enjoy computer vision. 6181