Measuring abstract reasoning in neural networks

Re:Measuring abstract reasoning in neural networks by stared on Sunday July 15, 2018 @07:58AM (#17534856)

Abstraction is looking at some properties, while abstracting others (removing, ignoring). Even simple artificial neural networks for computer vision can do so - e.g. ignoring noise, small rotations, reflections or the background.

If, or when, it gets to human-level (for a given task) is another question. But saying that computers cannot learn how to abstract data is demonstrably false.

Vide https://distill.pub/2017/feature-visualization/ for internal representations.

[ Parent | on HN ]

Re:Measuring abstract reasoning in neural networks by guntars on Sunday July 15, 2018 @11:39AM (#17535672)

I’d say neural networks abstract over successive layers until the answer becomes self evident. That’s not the same as reasoning though.

[ Parent | on HN ]

Re:Measuring abstract reasoning in neural networks by stared on Sunday July 15, 2018 @12:12PM (#17535814)

What do you exactly mean by "reasoning"?

[ Parent | on HN ]

Re:Measuring abstract reasoning in neural networks by ovi256 on Tuesday July 17, 2018 @04:07PM (#17552823)

Reasoning is whatever is left after you take away everything neural networks do. Obviously!

[ Parent | on HN ]