Gary Marcus

@GaryMarcus In your paper "DL: A Critical Appraisal" you mention a test you did for generalization during 1998. Its is still a good generalization test? Which is in your opinion the best generalization test for a AI?

There are (and should be) many different ways to assess generalization, but you might start with @LakeBrenden’s updates to my late 90’s work. And look at @dileeplearning’s work on how DQN fails in tiny changes to Atari games.

Do you consider at Vicarious that your tech & team have achieved generalization?

Of course not. We have good generalization in some limited settings. We think we have some of the right ingredients, but a lot more work remains.

@MelMitchell1 compare the above to the screenshot from another company I shared recently ... just sayin’.

Yup, I'm a big fan of @dileeplearning and @vicariousai Just to clarify from my previous posts: I wasn't saying that DeepMind doesn't hype their work; they absolutely do. I was just questioning your criticism that they don't have a successful deep RL *commercial product*.