How good are humans at zero-shot identification of dogs?
It is estimated that there are 900 million dogs in the world. That is a lot of dogs, and we assure you that each one is a good dog \cite{dog-rates}. Unfortunately, we don’t have enough time to meet each dog de novo, and have inevitably needed to rely on word-of-mouth to learn about which dogs to meet and when. In this work, we demonstrate human-level performance for zero-shot dog recognition from features described by other humans. Human performance is robust (>85\% accuracy), even when presented with challenging comparisons. This accuracy is in the same ballpark as Karpathy et al.’s work on a human baseline for ImageNet \cite{karpathy}. We believe that this work will help future researchers develop AI-based tools for super-human performance on word-of-mouth-based human-dog introductions. From a neuroscience perspective, this work also establishes the presence of a seeming information barrier between the visual cortex and the language system of the human brain.