WEBVTT - The Complicated Story of an AI Artist 0:00:04.120 --> 0:00:07.200 Get in touch with technology with tech Stuff from half 0:00:07.200 --> 0:00:13.680 stuff works dot com. Hey there, and welcome to tech Stuff. 0:00:13.720 --> 0:00:16.480 I'm your host, Jonathan Strickland. I'm an executive producer with 0:00:16.520 --> 0:00:19.639 how Stuff Works in a love all things tech, and 0:00:19.680 --> 0:00:23.920 today we're going to tackle a story that recently unfolded recently. 0:00:24.000 --> 0:00:26.480 As of the recording of this show, I'm sitting in 0:00:26.560 --> 0:00:30.639 the recording studio on October two, thousand eighteen. It's not 0:00:30.680 --> 0:00:33.159 my normal studio either, So if you hear other noises, 0:00:33.360 --> 0:00:36.240 that's because we've got noisy people walking around the office 0:00:36.240 --> 0:00:39.400 and I'm in a different studio. That's commentary. But this 0:00:39.479 --> 0:00:44.120 story unfolded just at the very end of October. That 0:00:44.280 --> 0:00:48.440 was when the auction house Christie's, put a special item 0:00:48.560 --> 0:00:52.200 up on the auctioning block. It was a somewhat blurry 0:00:52.280 --> 0:00:56.480 portrait of a man dressed in antiquated clothing. It looked 0:00:56.520 --> 0:00:58.560 like a painting that could have come from the eighteenth 0:00:58.680 --> 0:01:02.240 century from one of any number of artists, but it 0:01:02.360 --> 0:01:06.279 was in fact a much more recent painting. The artist 0:01:06.360 --> 0:01:09.880 was not a famous painter. In fact, the artist wasn't 0:01:10.000 --> 0:01:14.600 a person. It was an artificially intelligent algorithm that created 0:01:14.640 --> 0:01:18.800 the portrait through the process of machine learning. And what's more, 0:01:19.280 --> 0:01:23.360 the group of human artists who supplied the AI generated 0:01:23.400 --> 0:01:27.440 portrait had taken a great deal of direction, let's say, 0:01:27.560 --> 0:01:31.280 from a different computer programmer, but perhaps did not do 0:01:31.720 --> 0:01:36.040 as much to attribute that coder's work to the creation 0:01:36.080 --> 0:01:39.039 of this portrait that they should have done. So what 0:01:39.160 --> 0:01:41.919 we have here sounds a bit like a twenty first 0:01:42.040 --> 0:01:45.800 century futuristic art heist, only this isn't about stealing a 0:01:45.840 --> 0:01:50.440 work of art, but rather a means of generating art itself, 0:01:50.720 --> 0:01:54.240 and it's creating a lot of interesting conversations about concepts, 0:01:54.360 --> 0:01:57.600 ranging from what is art in the first place, to 0:01:57.680 --> 0:02:00.960 the practical applications of machine learning to the nature of 0:02:01.040 --> 0:02:04.840 open source code. So let's dive down into this, because 0:02:04.880 --> 0:02:08.520 when it comes to discussing our how technology interacts with 0:02:08.520 --> 0:02:11.720 our lives, this is a doozy of a story. It 0:02:11.800 --> 0:02:15.000 highlights not just technological issues but human ones that just 0:02:15.120 --> 0:02:18.880 happened to intersect with technology. So to begin with, let's 0:02:18.919 --> 0:02:22.440 talk about the tech behind generating this portrait in the 0:02:22.480 --> 0:02:27.119 first place. It is an application of machine learning. That's 0:02:27.120 --> 0:02:29.880 one of those topics we've talked about a lot on 0:02:30.040 --> 0:02:34.960 tech stuff, especially recently. But basically, machine learning is all 0:02:35.000 --> 0:02:38.920 about designing processes that allow machines to parse data in 0:02:38.960 --> 0:02:42.480 some useful way and then apply the results of those 0:02:42.520 --> 0:02:46.280 operations to future problems. But that's pretty darn vague, right, 0:02:46.360 --> 0:02:48.640 that's not that doesn't really tell you anything useful if 0:02:48.639 --> 0:02:51.280 you dive down a bit further, it's about creating a 0:02:51.320 --> 0:02:55.160 framework within which machines can learn to perform a task 0:02:55.600 --> 0:02:58.800 without having to be programmed to do it. So let's 0:02:58.880 --> 0:03:01.280 use an example, and it's one I've talked about a 0:03:01.280 --> 0:03:03.600 lot because it was one of the early examples of 0:03:03.600 --> 0:03:06.240 what machine learning could do once it reached a certain 0:03:06.320 --> 0:03:10.280 level of sophistication. Back in two thousand twelve, Google showed 0:03:10.320 --> 0:03:14.799 how their computer scientists teams had taught an AI algorithm 0:03:14.919 --> 0:03:19.840 or neural network to recognize images of cats. Now, this 0:03:19.919 --> 0:03:21.920 was perhaps a funny way of showing an approach to 0:03:21.960 --> 0:03:25.200 a difficult problem. So if you want a computer to 0:03:25.280 --> 0:03:28.720 recognize an image of a cat, if it's a specific 0:03:28.760 --> 0:03:31.079 image of a cat, you have a couple of different options. 0:03:31.360 --> 0:03:34.280 One is, you can program the computer so that when 0:03:34.320 --> 0:03:38.600 it encounters a specific arrangement of pixels for this particular image, 0:03:38.880 --> 0:03:41.880 it recognizes that as the image of a cat, and 0:03:41.920 --> 0:03:45.440 that you have programmed the computer to say, when you 0:03:45.840 --> 0:03:50.080 see this arrangement of pixels, then that means this is 0:03:50.200 --> 0:03:52.720 a cat. The computer doesn't understand what a cat is, 0:03:52.800 --> 0:03:56.400 it doesn't have any context. It doesn't understand what any 0:03:56.440 --> 0:03:59.280 other picture of a cat might be because that would 0:03:59.320 --> 0:04:03.600 be a different arrangement of pixels. So you could program 0:04:03.600 --> 0:04:05.480 a computer to do this and it would be able 0:04:05.520 --> 0:04:07.880 to do it with that one image. But if you 0:04:07.880 --> 0:04:09.680 gave it a different image of a cat, or even 0:04:09.680 --> 0:04:12.320 an image of the same cat, but it's a different picture, 0:04:12.880 --> 0:04:14.960 the computer would not be able to identify it. You 0:04:14.960 --> 0:04:18.560 would have to repeat the entire process from beginning to 0:04:18.720 --> 0:04:21.400 end to get the same result. And once you start 0:04:21.400 --> 0:04:25.440 adding up images, you realize this is not really an 0:04:25.440 --> 0:04:30.680 efficient means of teaching a computer anything. Or you could 0:04:30.680 --> 0:04:34.560 create an artificial neural network that examines the pixels in 0:04:34.600 --> 0:04:37.719 an image, and each neuron might be looking at a 0:04:37.760 --> 0:04:40.440 different element of the data to determine if that data 0:04:40.600 --> 0:04:44.200 was consistent with images of cat pictures. So we've talked 0:04:44.200 --> 0:04:48.160 about this recently too, and artificial neuron can take in 0:04:48.320 --> 0:04:53.080 multiple bin binary points of data's euros and ones and 0:04:53.080 --> 0:04:56.920 then create a single binary output. So it might be 0:04:56.920 --> 0:05:00.840 looking at specific features that might have to do with ears, 0:05:00.880 --> 0:05:03.600 for example, and if it detects that the ears are 0:05:03.680 --> 0:05:06.840 consistent with those of a cat, it might pass a 0:05:06.880 --> 0:05:10.080 positive response further down the neural network, and a full 0:05:10.120 --> 0:05:12.839 collection of all these looking at multiple points of data 0:05:13.160 --> 0:05:17.320 would allow the computer to come to a decision does 0:05:17.400 --> 0:05:21.640 this image represent a cat or does it represent something else. So, 0:05:21.720 --> 0:05:26.120 in this way, by feeding thousands or tens of thousands 0:05:26.200 --> 0:05:29.320 or hundreds of thousands of images to a computer, you 0:05:29.360 --> 0:05:32.160 can train it to recognize cats. And the more you 0:05:32.240 --> 0:05:35.560 train it and the more closely you're able to tweak 0:05:35.720 --> 0:05:39.600 the network so that it waits certain elements more than others, 0:05:40.240 --> 0:05:43.640 the better it gets. So the tweaking makes the network 0:05:43.680 --> 0:05:47.400 more capable and eventually get to a point where it 0:05:47.480 --> 0:05:50.840 can identify a picture as either being a cat or 0:05:50.920 --> 0:05:56.080 not a cat with pretty good results. Um Back in 0:05:56.120 --> 0:05:59.719 two thousand twelve when Google was talking about this, it 0:05:59.839 --> 0:06:03.720 was still a little jankie. It could sometimes recognize a cat, 0:06:04.000 --> 0:06:06.479 and sometimes it would think that a person was a 0:06:06.480 --> 0:06:09.240 cat or that a cat was a person, So it 0:06:09.400 --> 0:06:13.799 was not infallible, but it was pretty good. Now, because 0:06:13.839 --> 0:06:16.920 I've covered artificial neural networks in recent episodes of tech Stuff, 0:06:17.240 --> 0:06:20.119 I'm not gonna go through the whole thing all over again. 0:06:20.160 --> 0:06:22.279 That high level I just gave you that's a pretty 0:06:22.279 --> 0:06:24.919 good starting point. It's just important to remember that the 0:06:25.000 --> 0:06:28.560 general output here is through training and network using that 0:06:29.120 --> 0:06:32.000 input data set in this case or in the case 0:06:32.040 --> 0:06:35.080 of that example, hundreds of thousands of images of cats. 0:06:36.000 --> 0:06:40.400 Machine learning can actually take a few different approaches. The 0:06:40.440 --> 0:06:44.120 one that I sort of outlined earlier would kind of 0:06:44.160 --> 0:06:48.040 fall into the category of supervised machine learning. See in 0:06:48.120 --> 0:06:50.880 that approach, we human beings are trying to teach a 0:06:50.920 --> 0:06:56.640 machine through algorithms and data sets two recognize something that 0:06:56.680 --> 0:07:00.000 we already know the answer for. Right, you can look 0:07:00.000 --> 0:07:02.640 get a picture, and you can recognize whether that picture 0:07:02.680 --> 0:07:05.120 is of a cat or not, so you already know 0:07:05.200 --> 0:07:07.320 the answer. You're not asking the computer to give you 0:07:07.400 --> 0:07:10.200 new information. You're trying to teach the computer to do 0:07:10.280 --> 0:07:16.360 something that you already can do. So we human beings 0:07:16.440 --> 0:07:20.160 are able to supervise the machine as it is learning 0:07:20.200 --> 0:07:23.560 this process and make those minor adjust adjustments that are 0:07:23.600 --> 0:07:26.160 needed throughout the system in order for it to get 0:07:26.200 --> 0:07:29.920 better at its job. That is supervised machine learning. We 0:07:29.960 --> 0:07:32.280 can keep working with it until it reaches what we 0:07:32.320 --> 0:07:36.080 consider to be an acceptable level of success, which doesn't 0:07:36.080 --> 0:07:37.480 mean it has to be perfect. It just has to 0:07:37.480 --> 0:07:39.840 be good enough for whatever it is we're building it for. 0:07:40.480 --> 0:07:46.160 But there's another approach called unsupervised machine learning, and as 0:07:46.400 --> 0:07:50.040 you might imagine, this is different from the previous one. 0:07:50.160 --> 0:07:53.520 On this approach, you only have input data and your 0:07:53.520 --> 0:07:56.120 goal as a human is to learn more about that 0:07:56.240 --> 0:07:59.640 data itself. So you don't have a correct answer in mind. 0:08:00.040 --> 0:08:03.400 You don't already know that the data represents, say a 0:08:03.520 --> 0:08:06.360 cat in a photo. It's a different type of problem 0:08:06.400 --> 0:08:09.680 you're looking at. Uh. The machine is learning about the 0:08:09.760 --> 0:08:13.600 nature of the information itself, including how different points of 0:08:13.680 --> 0:08:17.360 data relate to one another or correspond with other data, 0:08:17.680 --> 0:08:21.080 and you in turn can learn more about the information 0:08:21.120 --> 0:08:24.000 as well. So within this category you have a couple 0:08:24.240 --> 0:08:29.040 of subcategories. There are clustering problems. With a clustering problem, 0:08:29.120 --> 0:08:32.800 you're learning about the groupings within data. So one example 0:08:32.880 --> 0:08:35.079 might be that you have a population of customers. Let's 0:08:35.080 --> 0:08:37.520 say you own a business. You've got customers. You have 0:08:37.600 --> 0:08:40.920 data that represents all these different customers, and you're using 0:08:40.920 --> 0:08:45.080 the collective behaviors of those customers to sort them into 0:08:45.160 --> 0:08:48.360 meaningful groups so that you can better serve each of 0:08:48.400 --> 0:08:52.600 those groups. Maybe you learn that there are four basic 0:08:52.679 --> 0:08:55.439 types of customers, and that helps you plan out your 0:08:55.440 --> 0:08:59.160 business so that you can cater it to those four types. 0:09:00.000 --> 0:09:03.280 But another type of problem in unsupervised machine learning is 0:09:03.280 --> 0:09:06.960 called an association problem. Now, in those problems, you want 0:09:07.000 --> 0:09:09.880 to learn rules that describe large parts of the data 0:09:09.960 --> 0:09:12.440 that you're feeding into the system. So, for example, let's 0:09:12.440 --> 0:09:15.000 go back to you run a business. You've got this 0:09:15.040 --> 0:09:17.880 big pool of customers, and you're feeding all the customer 0:09:18.120 --> 0:09:22.280 behavior data into your system. It might tell you that, hey, 0:09:22.520 --> 0:09:26.280 it turns out that of the customers who are buying 0:09:26.880 --> 0:09:30.840 widgets go on to buy sprockets. So that would tell you, hey, 0:09:31.000 --> 0:09:33.320 now I know more information. I know that if I 0:09:33.360 --> 0:09:35.480 sell a widget to someone, there's a good chance I 0:09:35.520 --> 0:09:38.600 can upsell that and include a Sprocket as well. So 0:09:38.640 --> 0:09:41.760 I'm going to tailor my business approach to try and 0:09:41.800 --> 0:09:44.440 take advantage of that. Now, the reason I went through 0:09:44.480 --> 0:09:47.240 all of this is to explain that the type of 0:09:47.320 --> 0:09:51.439 artificial intelligence algorithm that was used to produce the painting 0:09:51.559 --> 0:09:53.360 I was talking about at the top of the show, 0:09:53.920 --> 0:09:58.400 falls into a group called generative adversarial networks or g 0:09:58.559 --> 0:10:03.120 a N or a GAN. These are used in unsupervised 0:10:03.240 --> 0:10:07.400 machine learning applications. So it's in that second category I 0:10:07.480 --> 0:10:11.360 was just talking about. So what is with this name? 0:10:11.440 --> 0:10:17.720 What is a generative adversarial network? Well, for one thing, 0:10:18.360 --> 0:10:23.120 it actually uses a pair of deep neural net architecture networks. 0:10:23.600 --> 0:10:27.200 These two nets are in competition with one another. That's 0:10:27.200 --> 0:10:31.480 why it's called an adversarial network. You have these two 0:10:31.520 --> 0:10:37.800 different constructs that are working against each other. The approach 0:10:37.880 --> 0:10:40.840 was first proposed by researchers at the University of Montreal, 0:10:41.240 --> 0:10:44.559 and we chiefly associate the concept with a guy named 0:10:44.600 --> 0:10:49.560 Ian Goodfellow. Ian Goodfellow wrote the definitive paper on the 0:10:49.600 --> 0:10:53.559 subject back in two thousand and fourteen, and it is fascinating. 0:10:53.679 --> 0:10:56.480 So from a very high level, what's happening is that 0:10:57.160 --> 0:11:00.320 you have a neural network called the generator and you 0:11:00.360 --> 0:11:04.120 have a second year old network called the discriminator. So 0:11:04.280 --> 0:11:08.840 you're feeding the discriminator your input data. Let's again go 0:11:08.960 --> 0:11:12.880 with pictures of cats, So actual pictures of cats photographs 0:11:12.960 --> 0:11:16.040 of cats. If you will, you're you're feeding photographs of 0:11:16.080 --> 0:11:20.360 cats to the discriminator. The generator's job is to create 0:11:21.280 --> 0:11:26.120 a an image that fools the discriminator into thinking that 0:11:26.120 --> 0:11:29.520 that's a legitimate photograph of a cat, but in fact 0:11:29.600 --> 0:11:34.040 it was created or generated by the generator. So you've 0:11:34.040 --> 0:11:37.760 got two processes going on at the same time. The 0:11:37.840 --> 0:11:41.280 generator is trying to create essentially a forgery or a counterfeit. 0:11:41.880 --> 0:11:46.720 It's it's creating something from scratch to fool the discriminator 0:11:46.760 --> 0:11:50.880 into thinking this is a legitimate piece of data from 0:11:50.920 --> 0:11:55.320 the training data set. The discriminator is looking at each 0:11:55.360 --> 0:11:58.360 image and thinking, all right, now does this represent a 0:11:58.440 --> 0:12:01.800 real picture or is this something that is coming from 0:12:01.840 --> 0:12:04.800 the generator that's designed to fool me, And the two 0:12:04.880 --> 0:12:08.240 are working against each other. Both networks learn as this 0:12:08.280 --> 0:12:11.199 goes on. If the discriminator gets an image and rejects it, 0:12:11.720 --> 0:12:15.160 that becomes a feedback to the generator and the messages. Essentially, 0:12:15.800 --> 0:12:18.360 this was not good enough, and the generator starts to 0:12:18.800 --> 0:12:22.960 try again, taking a slightly different approach. If the discriminator 0:12:23.040 --> 0:12:25.840 accepts it, the generator says, ah ha, you're onto something. 0:12:26.160 --> 0:12:31.240 But then you can tweak the discriminator and say this 0:12:31.320 --> 0:12:33.440 was wrong. You you got this part wrong, and it 0:12:33.480 --> 0:12:36.000 can start to try and look for signs that might 0:12:36.080 --> 0:12:40.440 otherwise fool it. The goal here is that you are 0:12:40.480 --> 0:12:44.320 going to have a generator producing better and better versions 0:12:44.480 --> 0:12:48.440 of whatever it is you're trying to create. And that 0:12:48.520 --> 0:12:52.520 could be a picture, it could be text, it could 0:12:52.559 --> 0:12:56.400 be music. You could feed any sort of data to 0:12:56.559 --> 0:13:00.520 both of these systems in an effort to deuce a 0:13:00.600 --> 0:13:05.080 computer generated version of that thing, and as long as 0:13:05.080 --> 0:13:08.680 it reached a certain level of quality, the discriminator won't 0:13:08.679 --> 0:13:10.600 be able to tell the difference, and then you've got 0:13:10.640 --> 0:13:14.480 yourself a computer generated whatever it might be, in this case, 0:13:15.200 --> 0:13:18.480 a painting. I'll explain more about the specifics of this 0:13:18.559 --> 0:13:20.560 case in just a moment, but first let's take a 0:13:20.640 --> 0:13:30.640 quick break to thank our sponsor. So a couple of 0:13:30.720 --> 0:13:34.320 years ago, there were computer scientists at Microsoft as well 0:13:34.360 --> 0:13:38.200 as tu Deft University, and they were working together with 0:13:38.280 --> 0:13:41.080 a banking company I n G to create a brand 0:13:41.080 --> 0:13:45.440 new painting in the style of the painter Rembrandt. This 0:13:45.559 --> 0:13:49.640 project involved processing high resolution digital scans of three hundred 0:13:49.840 --> 0:13:56.320 forty six different images of Rembrandt's works, specifically portraits of men. 0:13:56.840 --> 0:14:00.480 That information was fed to a deep learning algorithm that 0:14:00.600 --> 0:14:05.560 analyzed Rembrandt's style and also the techniques that were common 0:14:05.640 --> 0:14:08.160 across all the images. What were the common elements that 0:14:08.200 --> 0:14:12.600 were found in those numerous paintings, And eventually this machine 0:14:12.679 --> 0:14:15.480 was told, or this system was told to produce a 0:14:15.600 --> 0:14:20.680 new painting based on those uh those common factors. And 0:14:20.720 --> 0:14:23.160 so it narrowed down the approach to be a portrait 0:14:23.480 --> 0:14:26.320 of a Caucasian white male because that's what most of 0:14:26.360 --> 0:14:30.760 Rembrandt's portraits were of, somewhere between the ages of thirty 0:14:30.760 --> 0:14:33.720 and forty, wearing white and black clothing, because again that 0:14:33.800 --> 0:14:37.600 was the vast majority of the portraits that Rembrandt created, 0:14:37.880 --> 0:14:42.640 and the focus of the subject was off to the right, 0:14:42.720 --> 0:14:46.280 like looking slightly off to the right, because a lot 0:14:46.280 --> 0:14:48.480 of the subjects in the other paintings were doing the same. 0:14:49.040 --> 0:14:52.520 The algorithm also analyzed the faces of all those portraits 0:14:52.520 --> 0:14:54.560 and came up was sort of a kind of a 0:14:54.640 --> 0:14:57.600 mishmash average of them to produce the face of the 0:14:57.640 --> 0:15:00.840 fictional Dutch gentleman in the new painting. To go a 0:15:00.880 --> 0:15:04.400 step further, the team then added depth to this painting. 0:15:04.440 --> 0:15:06.800 It was a two dimensional image, and then they decided 0:15:06.840 --> 0:15:09.000 to add some depth. They included some ridges and some 0:15:09.080 --> 0:15:13.080 bumps that would have been created from brush strokes onto 0:15:13.240 --> 0:15:17.320 a two dimensional surface. So if you're using paint, then 0:15:17.560 --> 0:15:19.800 it's actually a three dimensional image. You know, if you 0:15:19.840 --> 0:15:23.120 get super close enough, you can see raised areas and 0:15:23.560 --> 0:15:26.680 dips and trenches and stuff like that that the brush leaves. 0:15:26.800 --> 0:15:31.560 And it all depends upon your painting technique how these 0:15:31.600 --> 0:15:35.280 get laid out on canvas. So the team added those 0:15:35.320 --> 0:15:39.640 details in to make it look even more authentic. Ultimately, 0:15:39.720 --> 0:15:43.640 the design was printed using thirteen layers of ultra violet 0:15:43.720 --> 0:15:46.840 based inc and the result is a work that looks 0:15:46.880 --> 0:15:49.600 like it could have come from Rembrandt, complete with techniques 0:15:49.600 --> 0:15:53.760 Rembrandt used in actually making his brushstrokes. And that's just 0:15:53.880 --> 0:15:57.480 one high profile example of computers generating paintings after being 0:15:57.520 --> 0:16:01.040 fed information about works that human artists have created. Now, 0:16:01.040 --> 0:16:05.040 as get back to the story of the recently auctioned painting. Now, 0:16:05.600 --> 0:16:07.440 to do that, we have to talk about a young 0:16:07.480 --> 0:16:12.560 man named Robbie Barrett. Barrett is nineteen years old and 0:16:12.680 --> 0:16:16.200 is attending Stanford and has been doing some really interesting 0:16:16.240 --> 0:16:19.840 work in machine learning. It was his code that would 0:16:19.840 --> 0:16:22.640 be the basis for the computer generated portrait that was 0:16:22.760 --> 0:16:26.040 recently auctioned off. Barrett's work was going a step further 0:16:26.560 --> 0:16:30.800 than copying the style of an established artist. Barrett's algorithms 0:16:30.920 --> 0:16:34.640 would work to create new images after having analyzed numerous 0:16:34.720 --> 0:16:38.120 real world examples. So just a couple of years ago, 0:16:38.480 --> 0:16:42.000 the state of the art in GAN networks or GN 0:16:42.040 --> 0:16:46.680 networks might produce some really disturbing images, like there are 0:16:46.720 --> 0:16:50.200 early pictures of GAN attempts at making realistic human faces 0:16:50.680 --> 0:16:53.960 that were not terribly successful, and that's because those networks 0:16:53.960 --> 0:16:57.560 were able to recognize certain basic visual elements and images, 0:16:58.160 --> 0:17:02.880 but not understand the reation ships between multiple elements within 0:17:02.960 --> 0:17:05.200 an image, so you could end up with a face 0:17:05.480 --> 0:17:11.040 with really extreme features like pronounced asymmetry. But over just 0:17:11.080 --> 0:17:13.040 a short amount of time, people have developed much more 0:17:13.040 --> 0:17:17.160 sophisticated GAN algorithms and performance has improved, and there of 0:17:17.160 --> 0:17:20.440 course artists who have gone in a different approach, specifically 0:17:21.240 --> 0:17:25.600 emphasizing some of these more absurd elements in order to 0:17:25.640 --> 0:17:29.920 get that kind of a result when you're actually producing art. 0:17:30.560 --> 0:17:33.439 Verrett created GAN algorithms that could generate all sorts of 0:17:33.440 --> 0:17:37.800 interesting images. He was enabling computers to make art themselves. 0:17:38.240 --> 0:17:41.760 And sure, these computers were learning to create art after 0:17:41.800 --> 0:17:45.679 being fed numerous paintings and images from human artists. But 0:17:45.800 --> 0:17:47.920 you could argue that if you want to become a 0:17:48.000 --> 0:17:50.639 human artist, you have to do the same thing. You 0:17:50.680 --> 0:17:53.240 have to study art that was created by other people. 0:17:53.440 --> 0:17:57.960 So computers are no different. The computers weren't replicating specific works, 0:17:57.960 --> 0:18:00.840 they weren't trying to make a copy. They were learning 0:18:01.160 --> 0:18:07.280 various styles. Barrett would frequently put these images and also 0:18:07.320 --> 0:18:10.439 the algorithms he used to create those images up on 0:18:10.560 --> 0:18:13.920 get hub for free and open source. He also had 0:18:14.560 --> 0:18:18.760 uh people download these and upload their own art, and 0:18:18.800 --> 0:18:21.720 it was all in the spirit of this open source community. 0:18:23.200 --> 0:18:25.439 This way, not only could people use the tools that 0:18:25.480 --> 0:18:28.399 Barrett had created, they could understand how those tools worked, 0:18:28.840 --> 0:18:31.440 and perhaps in the future they can make their own tools, 0:18:32.000 --> 0:18:36.640 tweaking the approach the Barrett had used, maybe making art 0:18:36.720 --> 0:18:41.639 that was even more indistinguishable from human art, or perhaps 0:18:41.640 --> 0:18:44.760 going in a totally different direction, making something truly new 0:18:44.760 --> 0:18:47.560 and alien. By the way, some of the images created 0:18:47.560 --> 0:18:51.320 by Barrett's algorithms are a little unsettling. They can be 0:18:51.440 --> 0:18:54.359 surreal and absurd, and some of them even come across 0:18:54.359 --> 0:18:58.000 a little sinister to me. But that's my own interpretation. 0:18:58.040 --> 0:18:59.919 I mean, that is what art is all about, is 0:19:00.040 --> 0:19:02.520 the interpretation of the person looking at art. But they 0:19:02.560 --> 0:19:05.679 remind me of some of the horror movie effects you 0:19:05.760 --> 0:19:08.639 might see where the visual effects artists will distort a 0:19:08.680 --> 0:19:10.840 person's face for the effect of horror, like in the 0:19:10.880 --> 0:19:16.000 movie The Ring. Anyway, Barrett created several GAN algorithms and 0:19:16.040 --> 0:19:18.600 put them up online for others to use, and this 0:19:18.720 --> 0:19:21.520 in itself was not unusual. There are many in the 0:19:21.560 --> 0:19:24.800 digital art field who work on AI who have done 0:19:24.840 --> 0:19:29.760 similar things. Now he creates this code, Let's take a 0:19:29.800 --> 0:19:33.160 trip across the world from Stanford over to France. That's 0:19:33.160 --> 0:19:37.000 where three artists in their mid twenties were working in 0:19:37.040 --> 0:19:40.920 a group they had called Obvious and their stated goal 0:19:41.119 --> 0:19:45.280 is to promote ganism, that is, the art that has 0:19:45.320 --> 0:19:50.040 been generated through AI algorithms running on this GAN approach. Now, 0:19:50.080 --> 0:19:53.159 according to an article on Medium written by one of 0:19:53.200 --> 0:19:57.520 these artists, they quote want to send out an update 0:19:57.640 --> 0:20:00.600 of the state of the research and AI end quote. 0:20:01.200 --> 0:20:03.879 They want to do this they want to tell the 0:20:03.880 --> 0:20:06.560 world what is going on in the world of AI 0:20:06.680 --> 0:20:10.040 research through showing off artwork made by AI, so kind 0:20:10.040 --> 0:20:14.159 of a creative artistic way of talking about artificial intelligence. 0:20:14.960 --> 0:20:18.000 The group says that the value of the art may 0:20:18.040 --> 0:20:21.280 not be in the art itself, but rather the discussions 0:20:21.359 --> 0:20:25.040 that the art inspires, like what is it that makes 0:20:25.240 --> 0:20:30.720 art art? Can machines be creative? Who ultimately would you 0:20:30.800 --> 0:20:33.199 say is the artist in a work that was created 0:20:33.240 --> 0:20:37.160 by a machine? What does that art mean? Who does 0:20:37.160 --> 0:20:40.640 it belong to? That's a big one. So the artists 0:20:40.720 --> 0:20:44.200 reached out to Barrett when they were tackling this project. 0:20:44.560 --> 0:20:47.800 They wanted to use a gain algorithm to generate a 0:20:47.840 --> 0:20:50.480 portrait in a style similar to what you see in 0:20:50.600 --> 0:20:54.120 eighteenth century paintings out of Europe. The students have made 0:20:54.119 --> 0:20:56.720 it clear that Barrett had been a big part of 0:20:56.760 --> 0:20:59.880 their inspiration. More on that in just a second now. 0:21:00.080 --> 0:21:03.440 Members of Obvious began using gan code to generate portraits, 0:21:03.840 --> 0:21:06.760 and they created several of them, eleven in fact of 0:21:06.800 --> 0:21:14.119 a fictional noble family they named the Bellamy family B. E. L. A. M. Y. 0:21:14.280 --> 0:21:16.600 The name Bellamy itself was a bit of a pun 0:21:16.720 --> 0:21:19.919 and a reference to Ian Goodfellow, the guy who wrote 0:21:19.960 --> 0:21:23.520 that main paper on gangs. In the first place, Bellamy 0:21:23.680 --> 0:21:27.920 can be broken down into bell and Amy. That would 0:21:27.920 --> 0:21:30.679 mean all the different spellings. It would mean good friend 0:21:30.880 --> 0:21:34.120 or good fellow, which is kind of cute. Right. Well, 0:21:34.119 --> 0:21:38.320 the artists produced these portraits, and they are all of 0:21:38.440 --> 0:21:42.680 hollow eyed nobles that will stare right into the void 0:21:42.720 --> 0:21:45.920 in a way that actually that's getting off track. Never 0:21:45.960 --> 0:21:48.440 mind it. It creates me out a little bit. But 0:21:48.520 --> 0:21:51.600 the last in the line of portraits would be Edmund 0:21:51.840 --> 0:21:55.480 do Bellamy, the fictional noble whose portrait would go up 0:21:55.520 --> 0:22:00.640 on auction in October and fetched way more money than 0:22:00.920 --> 0:22:05.160 was anticipated and so obvious had fed to the algorithms 0:22:05.600 --> 0:22:09.520 numerous paintings from the eighteenth century to guide its efforts, 0:22:10.160 --> 0:22:13.160 and once they started producing these, they had each one 0:22:13.240 --> 0:22:16.680 signed with a line of code referencing the algorithm. They 0:22:16.760 --> 0:22:20.600 framed the machine generated portraits in golden frames, and when 0:22:20.760 --> 0:22:23.720 Edmund de Bellamy went up for auction, the best guess 0:22:23.720 --> 0:22:26.320 was that it would probably fetch between seven thousand and 0:22:26.359 --> 0:22:31.040 eleven thousand dollars. Instead, the winning bid was for more 0:22:31.119 --> 0:22:37.199 than four hundred thirty thousand dollars. So that raises a 0:22:37.280 --> 0:22:41.400 good question who the heck should get that money. Who 0:22:41.600 --> 0:22:46.840 was responsible for this painting and that would become something 0:22:46.960 --> 0:22:49.480 of a controversy. I'll explain more in just a second, 0:22:49.520 --> 0:22:52.600 but first let's take another quick break to thank our sponsor. 0:23:00.560 --> 0:23:04.639 So as the group Obvious was getting press coverage for 0:23:04.720 --> 0:23:08.040 the AI produced Bellamy portraits, this is before they had 0:23:08.119 --> 0:23:11.720 even put one up for auction, some people, including Barratt, 0:23:12.920 --> 0:23:17.239 express some disappointment with the group. They said that it 0:23:17.280 --> 0:23:21.480 looked like they had used Barrett's code to produce these portraits, 0:23:21.520 --> 0:23:24.440 and yet they weren't quick to attribute him. They didn't 0:23:24.440 --> 0:23:29.560 give him credit, at least not readily and not visibly 0:23:29.760 --> 0:23:33.440 in a lot of locations. And so his code, while 0:23:33.440 --> 0:23:36.920 it was open source and he didn't begrudge anyone from 0:23:37.119 --> 0:23:40.240 being able to use it, would have usually meant that 0:23:40.240 --> 0:23:44.360 people would give him credit. Typically in the open source community, 0:23:44.359 --> 0:23:47.679 it's considered bad form or even ghosh if you prefer 0:23:48.040 --> 0:23:52.080 to not give credit where credit is due. As to 0:23:52.119 --> 0:23:55.880 how much of the code was actually used unaltered, that 0:23:56.200 --> 0:23:58.840 is a bit of an open question. The artists that 0:23:58.920 --> 0:24:01.600 Obvious have admitted that they did use his code and 0:24:01.640 --> 0:24:05.520 they changed it a little bit. Some other artists say 0:24:05.520 --> 0:24:09.560 they believe that or more of the code was unaltered. 0:24:10.200 --> 0:24:13.200 One such artist, a New Zealander named Tom White, said 0:24:13.240 --> 0:24:17.280 he downloaded Barrett's code and ran it unaltered to see 0:24:17.280 --> 0:24:20.640 if he could produce images similar to those that Obvious 0:24:20.680 --> 0:24:24.439 had generated, and he said they look pretty close. So 0:24:24.480 --> 0:24:26.320 I took a look at as well. I would say 0:24:26.400 --> 0:24:29.760 that the ones that that White had created with that 0:24:29.880 --> 0:24:33.160 AI have a little bit more of the weird facial 0:24:33.240 --> 0:24:35.520 distortion thing going on than the ones that were made 0:24:35.560 --> 0:24:41.080 by Obvious, but they are fairly similar. Throughout the project, 0:24:41.440 --> 0:24:44.280