How does this work?
Like most image-search interfaces, you should use the text box below to type a description of what you want to see.
Typical image-search algorithms try to match words from your query to textual descriptions of the images and return the closest matches. The DeViSE-search, however, actually sees each image in the collection, and returns results using its understanding of how visual features correspond to natural language.
As a result, the DeViSE-search doesn't need to look at any of the images' associated metadata when you search. All it sees are the images, just like you.