I listened to a lecture by Schmdhuber https://www.youtube.com/watch?v=JSNZA8jVcm4 He talked about a recurrent network of a million weights which learned a policy to drive a car in on a simulated race track. He said that it was really really difficult to evolve something that had a million parameters, in fact no one is able to do that. So what they did was to develop compressed forms of the image source that the network was going to learn from.
I have no personal experience with deep learning networks but I was thinking that the problem with unsupervised deep learning is that without a policy governing the search space that it would use it would be similar to a programmed search space that acted in non-polynomial time. Even though a simple single level neural network is not in np, once you begin using it for unstructured deep learning the network may end up looking at variations of variations of relationships that it had previously adapted for. If these variants are not polynomial-time-effective, it is likely to go down into non-polynomial-time learning paths. Jim Bromer ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
