I've seen this talk in person last week and there are some new developments, in particular the coming up paper by EA Platanios and Mitchell 2014 where they desc

I've seen this talk in person last week and there are some new developments, in particular the coming up paper by EA Platanios and Mitchell 2014 where they describe a new way to reconcile consistency with correctness, i.e. how do you resolve conflicts between internally consistent but factually incorrect beliefs, and train the machine to learn from its mistakes (currently they just delete incorrect beliefs unlike human learners who incorporate their history of "fails" into making new decisions). Besides other things, they work on introducing time scale into learning and integrating NEIL (Never ending image learner).