947018640bf36a2bb609d3557a285329-AuthorFeedback.pdf

We would like to thank all reviewers for taking the time to review our work and for providing thoughtful suggestions. "original space", as the success of deep feature learning attests. Our testbeds (Sec 3.2) do indeed introduce more deliberate nuisance variables Specific settings might indeed warrant other types of interventions for safety/practicality. R1 suggests, where useful states may be discarded because of high agreement among bad hypotheses. MountainCar), but runs into optimization difficulties as the state space grows.