it would be interesting to see if NNs could learn tell how many objects are in the image, and/or where in the image the objects are.

if the answer is no, then I don’t think that even feedback would help in the original problem..