I just published What does it mean for a machine to ‘understand’? https://link.medium.com/vvNZ4pks80 1/
Show this thread
-
Your intuition is correct that existing correlation-based ML methods are not going to get you deep understanding. But they can get you something usable, and perhaps they are part of a larger and more comprehensive approach
What about GPT-2 responses to prompts the math word problems I shared, often nearly random, yet perhaps very occasionally correct. Those cases are bit like a broken clock that is correct twice a day. = shallow understanding?
