I’d be interested to see the results of applying this guy’s research to the task of developing an AI for segmentation and edge detection:
Jeff Clune
http://jeffclune.com/
Particularly this paper:
http://arxiv.org/pdf/1207.2743v1.pdf
I watched the talk that you suggested yesterday and it was very cool stuff. I’ll mention here my comments that I made then.
We don’t actually do any “evolving” of our neural networks. They are all based on the clever design of the PhDs and Post-Docs in our lab. I posted links to some papers from our lab on the wiki. After a lot of trial and error, we’ve come up with networks that work really well for the problem domain that we are working with. Given more time/effort/training data, we think that we can improve them even further.
That having been said, we are open to new directions of research. Last year we ran a design challenge where we asked other researchers if their techniques to the problem of neural segmentation. We are hoping to run a similar challenge again this year. We’d be happy to have your man Jeff and anyone else with awesome ideas as to how to automate this to step up and join the challenge. May the best networks win!