Vladimir, On 4/24/08, Vladimir Nesov <[EMAIL PROTECTED]> wrote: > > On Thu, Apr 24, 2008 at 6:32 PM, Steve Richfield > <[EMAIL PROTECTED]> wrote: > > > > Of course, some situations are more prone to superstitious learning than > > others. Religious explanations are a beautiful example of this - where > > everything that defies present explanation is simply credited to God. How > > are you going to keep your future AGI from becoming a religious zealot? > > > > How do I keep myself from becoming a Napoleon?
So, what's wrong with that?! It doesn't follow. Get > some background in philosophy of science from Bayesian perspective on > theories - and 'superstitiousness' of learning will be sorted out. LOL! That line of thought sunk a ~decade of efforts on the Russian translator. There is the "rub": Bayesian approaches teach that everything is in the probabilities, and that if you properly compute the probabilities that hopefully when everything comes together at the end, that you will have probabilities that are very near 0 or 1 that result. However, when you compute what fraction of one bit that those probabilities represent, you find that the information drops VERY fast with the loss of just a few percentage points. I agree with you that you can't just consider something to be "true" or "false" based on a few observations, but you DO have to make binary decisions based on whatever it is that you do know. Those decisions may reflect an underlying "belief" that isn't truly held, but who cares what the neurons/transistors are doing when all externally observable evidence is that they DO hold sometimes superstitious "beliefs". When applied to making decisions as to semantic intent, the Russian translator still made enough errors that the results were unusable despite being ~95% correct. My take on this was that no one on their team had ever seen a GOOD translation, where the translation notes are typically several times the length of the translation itself. Unannotated translations are cute, but very nearly worthless unless you are just looking for the bathroom or ordering breakfast - things that if wrong are easily corrected. I was once in Saudi Arabia and had occation to talk with the member of the "Metawa", their religious police and Ministry of Religion (yes, they do have such a ministry), who was in charge of converting visiting Christians. Their primary approach was to dig into whether they truly 100.0000...% BELIEVED all of the stuff they read in the Bible. Of course not - people are Beyesian too. The Metawa would then present their own version of whatever it was that wasn't perfectly believed, in the hopes that the Muslim version would be more perfectly acceptable. In short - it was a battle of models, and any open-minded and objective evaluation, setting politics and religious belief aside, that Islam is a MUCH less bad (notice that I didn't say "better") way to run a society than is Christianity, as ~98% of children live with both of their parents, etc. In short, people who are willing to accept religious models and are looking for the best religious model have some really good ones from which to choose. Also, I fail to see how Beyesian approaches will disfavor religious beliefs, as religious beliefs are absolutely perfect explainers, and hence will "float to the top" of probabilities once you roll in observation errors that will reduce all other probabilities. Steve Richfield ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com