(Maybe this Q should be moved to general discussions?)
I've got a question which is about two things: deep learning and the Winograd schema.
For anyone that doesn't know the Winograd schema is a suggested replacement for the Turing test, so it basically sets out to, if someone supplies some software/system and says, "this is intellegent", how do we actually know if that's the case? That's what these tests aim to answer. And the Winograd test does it by giving loads of these style of questions to the system:
"The city councilmen refused the demonstrators a permit because they feared violence. Who feared violence?"
So there's a situation description involving two entities, then a reference to one of those two entities using a pronoun (they, it, he, ... ) and then the question is, who/what is that pronoun referring to. In order to answer common and general sense is required, and an actual understanding of the situation, the interactions.
So my question, I don't know that much about deep learning but from what I do, it seems to be mainly about giving the system the end goal(s), examples of, and then also access to the necessary stuff/context/information/system, and then the deep learning system goes to work to work out how to achieve the end goal within the given system/situation.
Would deep learning systems be able to solve Winograd problems? Bearing in mind the textual descriptions in the Winograd questions can be about anything, so would potentially cover all human general knowledge of life and everything.
If I'm basically right about how deep learning operates, clearly it's no good for creativity. It's not going to come up with creative new ideas. That's simply not how it works. You give it the end answer, it goes to work to work out a way of achieving that end answer. It might come up with a creative route to the specified goal, but it's not going to come up with new interesting end results. The Winograd answers aren't exactly in the categroy of creative new ideas, but on the other hand the end goals are as many as the questions. It seems to me the answers to the Winograd questions are somewhere between creative new answers and speficied end goals. They're neither, they're inbetween those two things.
Discuss.
Thanks.
