Another failure mode for deep learning, precisely anticipated by my 2001 book, The Algebraic Mind.https://www.newscientist.com/article/2198761-deepmind-created-a-maths-ai-that-can-add-up-to-6-but-gets-7-wrong/ …
-
I read the arxiv (I used screenshots), and they were very upfront about failure modes that need further work (see below). They seem interesting and meaningful. Kinda like thwarting adversarial image attacks. Here's my beef: I take issue with your immediate jump to criticize /1pic.twitter.com/nKPw9K5Ktv
this discussion would make more sense to you @bensprecher and @IntuitMachine if you read not only the arxiv but the relevant parts of 2001 book that it cites (with wrong date as 2003) on the first page; it’s clear from both of your remarks that you have yet not done so.