Monica, I agree with Sridhar; you use “understand” to imply that deep nets connect a picture with “woman” in the way comparable to what a human does. In fact it doesn’t. What deep learning*does* do is a better job than ever before at accumulating the foreground pixels in an image into an statistical abstraction (a “symbol”) that recognizes pictures of other women better than any tech ever has. But that deep net has no deeper knowledge about that mass of pixels than that. And no doubt you’d agree, that’s not really “understanding” women at all.