2015.05.08
http://engineering.flipboard.com/2015/05/scaling-convnets/
What's interesting to me is what the failure modes / "artifacts" will look like - it seems like they'd have the potential to be weird, and almost "conceptual" - like the "sun tanning beds" example; the zoom on what seem to be dots on a pillow, it almost feels like the algorithm didn't reproduce the actual dots, it generated an image of a smaller number of similar dots. To me that seems reminiscent of how our minds tend to encode things via summaries, and then construct more detailed "recreations" on demand. (Which I think is how dreams can be so vivid and real-feeling, yet generated from bits and pieces of neural sparks.)
Oh, I hadn't realize they were using Neural Nets to play Atari games:
To learn to talk only the necessary: it takes around two years to learn to speak but it takes a lifetime to learn to shut up.
I'm hearing a lot of 'destination, destination, destination' here. I need to hear a lot more 'journey, journey, journey' out of you.