Re:Machine learning is fundamentally conservative
by l0b0 on Saturday January 04, 2020 @03:07AM
(#21953605)
I think the point is rather that you can't get a more useful prediction by choosing a lower probability description unless you have AGI. Only an AGI could tell that you're not in the mood for "Hey" to be followed by "darling", and only a superhuman AGI could realistically compensate for human bias in data sets.
[ Parent
| on HN ]
Re:Machine learning is fundamentally conservative
by kitsuac on Saturday January 04, 2020 @03:17AM
(#21953643)
Without AGI there are still cases when the lower probability prediction will be better, and will lead to escaping a local minima. I'd argue that the potential benefits of calibrating that axis dynamically exist with or without AGI.