Why Robot Brains Need Symbols

nautil.us

5 points·mindgam3··4 comments

The author, Gary Marcus, makes a cogent argument for why we need both deep learning and symbol manipulation to build "true" AI. what I found convincing was the argument that deep learning models currently can't even generalize to novel instances outside the training space in the field of object recognition, which is their strong suit. Why would any rational person believe that they have the capability to generalize to novel instances for other types of problems outside of perceptual classification?

Honestly as a native speaker of three symbolic languages (english, french, and chess) the need for symbolic manipuation aka approaches from classical AI seems fairly intuitive, if not downright obvious. I don't understand why Yann LeCun and other ML bigwigs find this controversial in the slightest.

Perhaps someone with more practical experience building AI could enlighten me on this subject.

I don't understand why Yann LeCun and other ML bigwigs find this controversial

The problem is not with the idea of combining symbolic AI and DL, it's with people like Marcus who seem to be all talk, without any results to show. He's been talking about this for a while now, I think it's time for a demo.

Fair point. However one could argue that symbolic AI is a harder problem, perhaps awaiting a theoretical breakthrough. And that the pace of progress in this area has slowed at least in part due to ML hype sucking all the air out of the room (along with the best talent).

along with the best talent

"Best talent" can decide for themselves what they should work on. As long as they produce results, no one cares how they do it.

Reply on news.ycombinator.com