I don't think that is directly applicable to this case (definitely not to the level of calling the problem 'solved'). The link describes a technique for traini

I don't think that is directly applicable to this case (definitely not to the level of calling the problem 'solved'). The link describes a technique for training a neural net to detect a type of image, an '8' for example, in an image regardless of location. However, in this case, you only have one sample of the target. You could run a net to attempt to extract a list of features (children, bronze), and compare that to all your images, but I don't think that's how the bing image search is structured. That would be closer to searching for '4 children bronze park' (which it does a great job of finding such things; just nothing related to the target image). Also, when you use convolution in the way described in the article, you would lose relative position information (or at least, dampen that part of the signal) of each of the bronze statues; as the entire point is to not care about the position of the object.