Károly Zsolnai-Fehér explains how BigGANs generate photorealistic images by training much larger networks than earlier models. The Inception Score jumped from 50 to 166, closing the gap with real images at 233. The code runs on a powerful GPU and offers better control over what the AI creates.
Approximately 50.
Approximately 166.
Approximately 233.
Dear fellow scholars, this is two minute papers with approximately 150 episodes ago, we looked at DeepMind's amazing algorithm that was able to look at a database with images of birds and it could learn about them so much that we could provide a text description of an imaginary bird type and it would dream up new images of them. It was a truly breathtaking piece of work. And its main limitation was that it could only come up with coarse images. It didn't give us a lot of details. Later we talked about Nvidia's algorithm that started out with such a coarse image, but didn't stop there. It progressively recomputed this image many times, each time with more and more details. This was able to create imaginary celebrities with tons of detail. This new work offers a number of valuable improvements over the previous techniques. It can train bigger neural networks with even more parameters, create extremely detailed images with remarkable performance. So much so that if you have a reasonably powerful graphics card, you can run it yourself. Here the link is in the video description. Training these neural networks is also more stable than it used to be with previous techniques. As a result, it not only supports creating these absolutely beautiful images, but but also gives us the opportunity to exert artistic control on the outputs. I think this is super fun. I could play with this all day long. What's more, we can also interpolate between these images, which means that if we have desirable images A and B, it can compute intermediate images between them. And the challenging part is that these intermediate images shouldn't be some sort of average between the two, which would be gibberish, but they would have to be images that are meaningful and can stand on their own. Look at this. Flying colors. And now comes the best part. The results were measured in terms of their inception score. This inception score defines how recognizable and diverse these generated images are. And most importantly, both of these are codified in a mathematical manner to reduce the subjectivity of the evaluation. This score is not perfect by any means, but it typically correlates well with the scores given by humans. The best of the earlier works had an inception score of around 50. And hold on to your papers, because the score of this new technique is no less than 166. And if we would measure real images, they would score around 233. What an incredible leap in technology. And we are even being paid for creating and playing with such learning algorithms. What a time to be alive. A big thumbs up for the authors of the paper for providing quite a bit of information on failure cases as well. We also thank in Silico Medicine for supporting this video. They are using these amazing learning algorithms to create new molecules, identify new protein targets with the aim to cure diseases and aging itself. Make sure to check them out in the video description. They are our first sponsors and it's been such a joy to work with them. Thanks for watching and for your generous support and I'll see you next time.
