Gary Marcus

Agreed. It is worth noting, however, that scene comprehension is not a bounded task. People do not comprehend everything about a scene. So it is not easy to define what it means to "solve" scene comprehension.

You don't need to have vision solved. You stated OpenAI js "ducking some of the hardest problems like knowing what entities exist and how to segment them etc.", but that subset of vision (detection + segmentation of a discrete known number of object given sim access) is not hard

Not hard for supervised learning, or not hard for unsupervised/RL? And not hard because DOTA has simple graphics, or not hard period?

It's much, much worse than that. Unlike DNNs, the brain can instantly see a complex object it has never seen before, i.e., without a prior representation. Still, you, Marcus, insist on calling for changes to the current representationalist DL paradigm. This is pathetic, man.

Given what I wrote above, if anybody still thinks that deep learning has any role to play in the future of computer vision or AGI, he or she is either stupid or a charlatan. Which do you want to be, Marcus? Either way, you will be just a footnote in the history of AI.

Amen. And even (to paraphrase Rudy Guiliani), "object recognition" ≠ object recognition. That is, "object recognition" as in ImageNet has been "solved", sort of, but object recognition in real life is far from solved.

object recognition ≠ object comprehension