DeepMind's PixelCNN generates photorealistic images from text descriptions. The original method was slow, creating pixels one at a time, but a new approach parallelizes generation and achieves over 100x speedups by treating weakly correlated pixels independently.
Dear fellow scholars, this is two minute papers with this is one of those new absolutely insane papers from the Google DeepMind guys. You are going to see a follow up work to an algorithm that looks at a bunch of images and from that it automatically learns the concept of birds, human faces or coral reefs. So much so that we are able to write a new sentence and it will generate a new close to photorealistic image. From this written description, this network is capable of creating images that are significantly different than the ones it has been trained on. This already sounds like science fiction. Completely unreal. This work goes by the name pixalcnn. We'll discuss a follow up work to that in a moment. The downside of this method is that these images are generated pixel by pixel and many of these pixels depend on their neighborhood. For instance, if I start to draw one pixel of the beak of a bird, the neighboring pixels have to adhere to this constraint and have to be the continu continuation of the beak. Clearly these images have a lot of structure. This means that we cannot do this process in parallel, but create these new images one pixel at a time. This is an extremely slow and computationally expensive process and hence the original paper showed results with 32 x 32 and 64 x 64 images at most. As we process everything sequentially, the execution time of the algorithm scales linearly with the number of pixels we can generate. It is like a factory where there are a ton of assembly lines but only one person to run around and operate all of them. Here the goal was to start generating different regions of these images independently, but only in cases when these pixels are not strongly correlated. For instance, doing this with neighbors is a no go. This is possible, but extremely challenging and the paper contains details on how to select these pixels and when we can pretend them to be independent. And now feast your eyes upon these spectacular results. If we are looking for a yellow bird with a black head, orange eyes and an orange bill, we are going to see much more detailed images. The complexity of the new algorithm scales with the number of pixels not linearly, but in a logarithmic manner, which is basically the equivalent of winning the jackpot in terms of parallelization. And it often results in a more than 100 times speedup. This is a factory that's not run by one guy, but one that works properly. The lead author, Scott Reed, has also published some more amazing results on Twitter as well. In these examples we can see the evolution of the final image that is generated by the network. It is an amazing feeling to be a part of the future. And note that there is a ton of challenges with the idea. This is one of those typical cases when the idea is only the first step and execution is king. Make sure to have a look at the paper for more details. According to our regular schedule, we try our best to put out two videos every week. That's eight episodes a month. If you feel that eight of these episodes is worth a dollar for you, please consider supporting us on Patreon. This way we can create more elaborate episodes for you. The channel is growing at a remarkable rate and your support has been absolutely amazing. I am honored to have an audience like you fellow scholars. We are quite close to hitting our next milestone and this milestone will be about giving back more to the scientific community and empowering other research projects. I've put a link to our Patreon page with the details in the video description. Thanks for watching and for your generous support and I'll see you next time. Ra.
