Nothing that really addresses the problem at the dataset level. Which kind of makes sense since foundation models are very costly to train and biases are often an afterthought for a lot of companies. The recent papers I remember try to band-aid fix the problem by, e.g., shifting a model's activiations after it was trained.
But since de-biasing training data is a general Machine Learning problem, something like this paper - they're evaluating basic de-biasing techniques for the Eduacation sector - might be a good starting point: https://www.sciencedirect.com/science/article/pii/S0957417423008254
Have you seen any recent papers discussing AI bias based on training data and how it's addressed?
Nothing that really addresses the problem at the dataset level. Which kind of makes sense since foundation models are very costly to train and biases are often an afterthought for a lot of companies. The recent papers I remember try to band-aid fix the problem by, e.g., shifting a model's activiations after it was trained.
But since de-biasing training data is a general Machine Learning problem, something like this paper - they're evaluating basic de-biasing techniques for the Eduacation sector - might be a good starting point: https://www.sciencedirect.com/science/article/pii/S0957417423008254