The TLDR is that to work at scale we sacrifice some correctness in the dataset, and what's really being evaluated here is closer to some notion of human agreeme

The TLDR is that to work at scale we sacrifice some correctness in the dataset, and what's really being evaluated here is closer to some notion of human agreement. The way the data was collected was not through a 1000-way classification. Instead, we'd query things like "red fox" from a search engine, and then have turkers do the binary task of YES/NO to clean up the results. This is a messy, noisy process but we do our best. In particular, the turkers saying YES/NO to red fox are completely unaware of what other classes are in the dataset - everything they say YES to becomes the "red fox" class in imagenet, in the final discriminative task among the 1000 categories. Also note that if an image contains a whole bunch of things (e.g. see left-most image here http://karpathy.github.io/assets/ilsvrc2.png http://karpathy.github.io/assets/ilsvrc2.png), the turker during data collection would be correct in saying that there is in fact, e.g. a "ruler" in that image, but at test time when you're seeing everything in that image, you only have 5 possible predictions so you have to guess about what you think the "true" label is.