RT @MinqiJiang@twitter.com
SAMPLR allows adaptive curricula to focus more on difficult scenarios (e.g. “driving on ice”), without the final policies *overestimating* their probability. In other words, it robustifies agents to difficult settings, without making them overly conservative or optimistic.
: 
@rockt Cool. I wonder if you could use this in offline classification to resample unbalanced data without changing the class distribution that the model expects?