Domain adaptation problem
What is domain adaptation? Suppose you have a great dataset with tons of data and nice and accurate labels, but all of those images/audio doesn’t look like your actual data, for example your product should work in infrared. If you still want to take something more than pretrain from that fancy dataset than you have to solve problem of domain adaptation. The talk is covering current approaches for domain adaptation, but is not limited to it. I would try to answer what approaches are fruitful and what are not so great.