I've just happened upon a Google Tech Talk where Geoffrey Hinton presents his new approach to training neural networks by creating a model capable of regenerating its training data. He begins his demonstration with a typical character OCR example, feeding in various character images and watching as the network converges on the correct output. Around the 25 minute mark he does something very cool, essentially running the network in reverse - for a chosen character, the image (now an output rather than an input) shivers and twists until the character appears recognizably, continuing to morph, shudder, and briefly slip to similar shapes. For whatever reason I'm reminded of the death throes of the T-1000, or the opening credits of a hip late 80s television show. It's neat.
Also, it's a little funny that Google would disable embedding on one of their own videos.