Ha! I was wondering this exact thing. How could it identify a sofa if it never learned about one? It seems very contrived. At the same time, the example is kind of clear, but we are making neural nets that are emulating Pollock and Dali, not Monet. If you can dig it. The whole beauty is the overlapping interstitial matrix of weighted values that leads to these beautiful discoveries. To rank such a fractal-like algorithm on whether or not it predicts a label satisfactory to the average human is to mis-apply the elegance and potential of these mathematical wonders, in my humble opinion.