Many companies and organizations hold Christmas card competitions for their employees’ children: GMV is no exception. This year with a difference, however: GMV’s Big Data and Artificial Intelligence Department (BDA) decided, as a small experiment, to take part with an AI-generated Christmas card, obviously unbeknown to the jury. Would we be able to fool the jury into taking the fake card as a real child’s creation? Could the AI card even win the whole caboodle? And what lessons could we learn from the experiment?
Creating and training GrinchGAN
GrinchGAN is a GMV-developed system dreamed up to meet the above challenge. It is based on Generative Adversarial Networks (GANs), the main idea of which is get two neural nets to vie with each other for the upper hand. A generating net called G produces random-noise-based images which are then sent to another discriminating net called D, which then tries to tell if the image received is genuine or G-generated.
By repeating this process the two nets learn from each other, thus producing a generator capable of creating totally new, uncannily-realistic images without copying or mixing original training images. The discriminator’s ability to tell the real images from the generated images is also honed, forcing the generator in turn to up its game. Figure 1 sketches out the idea in outline.In order to carry out the experiment, training images were first downloaded from Google Images; a sample of about 800 trees proved to be sufficient (some examples can be seen at the top left of Figure 3). Once the GAN had been trained up, we managed to obtain sufficiently tree-like images of reasonable resolution. We then made a manual selection of the most promising images (top right of Figure 3) then, in turn, short listing the ones that seemed capable of giving the best results. One of the competition constraints was that the Christmas card had to include GMV’s logo, so we decided to draw this in by hand, using the Paint program.
The final step involved the Neural Style Transfer technique; it was this final touch that gave the result the top-quality, drawing-like appearance. Neural style transfer consists in applying the “style” of a reference image to a target image, while conserving the “content” of the target image. Figure 2 shows how this technique works. In our case we chose as reference an image from one of the previous year’s Christmas card competition (bottom center of Figure 3) and our tree as target. We thus separated the style of the reference (colors, textures) from the content (hands) and replaced the style in the target, maintaining the content (tree, logo). We thus obtained the Christmas card shown at bottom left of Figure 3.
And the winner is …
The experiment turned out to have an intriguing result; none of the judges suspected the image was fake, proving that our generated image could compete on equal terms with children’s drawings and is also indistinguishable to the human eye.
Some questions do remain to be answered, however.
- Firstly, the resolution, although not bad, is insufficient for filling an A4 sheet without forfeiting quality.
- Secondly, one of the jury comments was a request to see the original drawing. This was a credibility-limiting factor and a big problem because there was in fact, with this procedure, no possibility of reproducing an original sketch, as would indeed have been possible with other manual techniques like watercolor.
The good news is that, although the jury would have liked to see the original drawing, they did see something ineffably “different” in it. Either we didn’t quite get it right or a child’s drawing is still more capable of pleasing the eye than an AI-generated one. Despite the abovementioned limitations, however, the image came second in its age category, suggesting that, in the near future, Grinch might be able to wrench the Xmas card prize from GMV employees’ children with only a computer and a bit of patience.
Although we would have liked to win the competition, personal and professional ethics urged us to set ourselves certain limits. True it is that the jury was not informed of the AI input, but the competition manager was told, to keep her in the picture. Far from banning the experiment she was dead keen for it to go ahead. But at all times she was on hand in case the AI card turned out the winner in its age category and would therefore have had to be disqualified.
Prima facie this might seem to be an innocent experiment. But it does provide food for thought about the short- and medium-term future of artificial intelligence, which is now capable of producing increasingly realistic Deep Fakes, of swaying the electorate in political campaigns and even publishing AI-generated fake news. Does this mean AI needs to be curbed? Can any model whatsoever be published to the public at large? A fascinating discussion on this matter in the context of language models (GPT-2) can be checked out at: https://openai.com/blog/better-language-models/.
Author: Antón Makarov
Las opiniones vertidas por el autor son enteramente suyas y no siempre representan la opinión de GMV
The author’s views are entirely his own and may not reflect the views of GMV