Women also Snowboard
This is a famous and interesting paper. They identify a common problem in so-called “captioning” networks: namely, they can be right for the wrong reasons. In the photo, we see that a network guesed it was a man sitting at a computer; but it only spent time “looking” at the computer to work this out. In other words, a computer was strongly correlated with the photo being of “a man at the computer” in the training data. In this paper they introduce some techniques to deal with this problem. Basically, their idea is that we can penalise the network for thinking about gender when no gender information is present, and reward it for thinking about gender when it is apparent. Furthermore, their approach is generally useful for other models and situations.
We can expect more technical results in this area to be implemented alongside the social techniques (i.e. having more diverse people involved in the building of AI systems).
