Yann LeCun

I'll let my eminent and distinguished FAIR colleagues Ross Girshick and Larry Zitnick answer that one: "Image classification works much better today than it did just a few years ago, thanks to the clever application of deep learning techniques developed by the research community. Deep learning has more or less cracked the most basic form of visual perception: classifying the dominant object in an image when the object comes from a constrained set of object categories. However, despite this success, deep learning is still very far from human-level visual perception when these constraints are...