On Fri, Nov 8, 2013 at 7:57 PM, Piaget Modeler <[email protected]> wrote: > > The article is free. I think scribd is trying to make some money for itself. > Did not know that. Sign up for a scribd account and you should be able > to download it for free. or instead I can e-mail it to you personally if you > prefer. > Let me know which way you want to go.
Why don't you just put the PDF file on your website? > The goal is to have the infant develop into a more mature general > intelligence. Are you going to raise it like a parent, or do you have some way to automate or speed up the training? > Some compression may be done on the audio video streams but largely > those percepts will be represented internally as monads. The retina compresses a 10^10 bit per second video stream down to about 10^7 bits per second transmitted over the optic nerve. By the time our visual perception reaches long term episodic memory, it is compressed to about 5 bits per second. We know this is the case because of limits on our ability to recall images or to notice differences as measured in cognitive tests. But we have no idea how to do the later stages of this type of lossy compression. We do have a pretty good idea of what the retina and lower layers of the visual cortex are doing. We can use neural networks like Hawkins' HTM to model the pattern recognition capabilities we have observed in animal and human experiments. Presumably the higher layers are doing a similar thing, but with more complex patterns such as faces, words, and other familiar objects. But to model this, we need a human brain sized neural network (1 petaflop) and train it on years worth of video, like 10^9 frames with 10^7 pixels each (10 petabytes). Our most ambitious experiments, like Google's cat face recognizer, fall far short of what a toddler can see. And that required 3 days of training on 8000 CPU cores on 10^-4 as much data. So how do you propose doing that with monads? Sure, they are elegant, but do you expect a 10^6 speedup? And, of course, such simple, highly repetitive structures as we typically use cannot model our genetically programmed fear of heights, snakes, and spiders. How do you code that into an untrained vision system? > Humans are very complex, PAM-P2 will be less so. Are you expecting human level results? If not, then what results would you consider successful? -- -- Matt Mahoney, [email protected] ------------------------------------------- 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
