Gary Marcus

I recently challenged Hinton and @YLecun to stand by the idea expressed in these words from their 2015 Nature paper, by making a public bet. Neither would. I stand by my assertion that is precisely this that holds back deep learning, except brave work by people like @egrefen.https://twitter.com/MLWave/status/1065310317935304704 …

I’m flattered, but—while I don’t want to put words in Yann’s or Geoff’s respective mouths—I don’t think any one is trying to strike symbols and concepts from the repertoir of our machine learning methods, even within DL. It’s just that they may become implicit...

“Needed to replace”? “Aetherial symbols” (title of Hinton talk I linked earlier)? I am trying to pin down what they actually mean, but likening symbols to aether that must be replaced is not just saying they are unconscious, it’s saying they are misguided and unnecessary.

Hm, I haven’t seen or heard that talk, but I agree that this is what it sounds like. Nonetheless, I don’t believe there’s a clean cleft between symbolic systems and connectionism, and as such I think a lot of DL research reintroduces such concepts without calling them as such.

“The fathers of AI believed that...human reasoning must...[have] rules of inference containing variables that got bound to symbols...this belief is...incorrect as the belief that a lightwave can only travel through space by causing disturbances in the luminiferous aether” -Hinton