I haven't used the feature, but the way it's implemented feels overly complicated, especially for something like keyword search (and not similar-image-search).
If they only use the Top10 categories in their feature vector for the documents, why don't they store these categories as tags on each documented and use standard inverted-index searching and scoring. I know the vector will express how much "beach" a certain image is, but your user-supplied query doesn't have a notion of how "much" beach the user expects, so the output can be a simple list ranked using standard term search mechanisms. What am I missing?
Hi, I wrote this post. The retrieval stage amounts to storing the top 50 categories in the inverted index and searching for the top 10 from the query. The harder part is the ranking. There are approaches to the ranking that are more akin to the techniques you'd use for text document search -- think of the classifier scores as TF-IDF values, treat query-to-category matching as something like synonyms -- but to my mind they're more complicated than our approach, not less. We did experiment a bit with some of these ideas, but the results were worse.
Comments
I haven't used the feature, but the way it's implemented feels overly complicated, especially for something like keyword search (and not similar-image-search).
If they only use the Top10 categories in their feature vector for the documents, why don't they store these categories as tags on each documented and use standard inverted-index searching and scoring. I know the vector will express how much "beach" a certain image is, but your user-supplied query doesn't have a notion of how "much" beach the user expects, so the output can be a simple list ranked using standard term search mechanisms. What am I missing?
Hi, I wrote this post. The retrieval stage amounts to storing the top 50 categories in the inverted index and searching for the top 10 from the query. The harder part is the ranking. There are approaches to the ranking that are more akin to the techniques you'd use for text document search -- think of the classifier scores as TF-IDF values, treat query-to-category matching as something like synonyms -- but to my mind they're more complicated than our approach, not less. We did experiment a bit with some of these ideas, but the results were worse.
Maybe the engineers were up for promo and needed something cool to stand out and inspire a blog post.