Pedro Domingos

It did (Werbos, Bryson & Ho, etc.), but he didn't know about it.

See this responsehttps://twitter.com/ylecun/status/1449006954886701069?t=eLh9lU6RBAiQff9bMd1QhA&s=19 …

So you don’t think something exists until it’s been shown to work. More accurate would be to give everyone credit for their part instead of just saying backprop was invented in 1985 (and even then by multiple separate people at once).

With this reasoning then we should credit Newton for inventing backprop. I mean Newton (/Leibniz) invented differentiation, so we should credit them for inventing the chains rule, which is apparently backprop too?

We should credit them with taking an important step in the invention of backprop, certainly.

Stop this Euro-centrism, Pedro! We all know that Brahmagupta invented and documented Zero, and everything after that was just an infinitesimal detail..

Yann can correct me, but I think I've heard him joke that backprop should really be credited to Lebniz because he invented the chain rule. So pin that one on him :)

The bigger point here, which perhaps we can all agree on, is that the most important (& hardest, at the time) part of backprop is realizing that it's a valid way to train models with multiple optima even though it has no guarantee of finding the global one (cf. Minsky & Papert).