Bumped into this on the web, thought it very neat, and thought it worth mentioning here in this forum for those interested:
RAISR is a machine-learning up-sampler. (I want to say the magic words "Neural Net" but the paper seems very careful to call it a politically-correct "Machine Learning" and I haven't actually read it properly, but I presume its NN.) It is trained by downscaling an image and then learning how to upscale it back up to the original.
Then, in real life, Google is serving 1 billion down-sampled images per-week to Android G+ clients and on the device upscaling them again for presentation, saving 75% bandwidth.