I've always found Bayesian Inference to be a more "solid" theory than neural nets. I like the idea, but its just a mish-mash of words to me right now...
Looking it over, the words seem to be... in the correct order. :-) "Proper use of Prior" and Baye's theorem. Heh, I'm going to have to pull out some textbooks to remind myself what these words mean.
EDIT2: Are they... generating a Bayesian model automatically? That's some meta-level theory going on here. A Bayesian Model is a graph of priors (If A then B 50% of the time, if B then C 25% of the time, etc. etc.). You observe "C", and then the program seems to be trying to find the "If A->B" probability and "If B->C" probability.
Unlike neural nets which are "unobservable", the human can theoretically look at the list of priors and understand the model that the computer has generated. As I stated earlier: I've always found this branch of AI to be more interesting than neural nets.
The title pretty much reflects today people's focus on such a progress. Maybe we can train attention networks to simulate focus of different groups of people, and I'm sure that "MIT" & "AI" will end up with high weights for average people while "Bayesian inference" will be on the top for researchers' model.
Bayesian program synthesis is an approach to ML that's been slowly brewing since the early parts of this decade, and as a member of a "sister lab" to Vikash's, I'm excited to see their POPL paper getting this kind of popular attention.
I'll be even more excited if their techniques from this paper can actually be widely reapplied! BPS as a technique hasn't had its "ImageNet" moment in which a strong, nontrivial, real-world benchmark is identified and then solved.
EDIT: The link in the article doesn't work for me. Here's the link I used instead: https://dl.acm.org/citation.cfm?id=3290350
Looking it over, the words seem to be... in the correct order. :-) "Proper use of Prior" and Baye's theorem. Heh, I'm going to have to pull out some textbooks to remind myself what these words mean.
EDIT2: Are they... generating a Bayesian model automatically? That's some meta-level theory going on here. A Bayesian Model is a graph of priors (If A then B 50% of the time, if B then C 25% of the time, etc. etc.). You observe "C", and then the program seems to be trying to find the "If A->B" probability and "If B->C" probability.
Unlike neural nets which are "unobservable", the human can theoretically look at the list of priors and understand the model that the computer has generated. As I stated earlier: I've always found this branch of AI to be more interesting than neural nets.