We would like to thank you for your thorough evaluation, helpful suggestions, and comments. We trained our loss predictor for five crop areas. Compared to the 5-crop ensemble, choosing one transform by our method gives almost the same performance, and selecting the two transforms achieves even better performance with less computational cost. Figure 2: Comparison for the same GPS transforms on the clean ImageNet set using ResNet-50. We trained our loss predictor on the searched GPS policies to choose ones specific for each test instance.

Weaknesses: The main claim of the paper is the hypothesis that the noise in the non-greedy objective's paper is the reason why the greedy method can outperform it. However, I think that the empirical methodology is not strong enough to back this claim up, as the experiments are carried out on a single dataset, using a single network architecture and reported with a single performance metric. I think that the hypothesis would be much more clearly substantiated if the noise in the gradients were shown to be a consistent trend in various setting; I am afraid that, in the current state, the conclusion could be an anecdotical performance of the given setting. In addition, if I'm correct, RL models are prone to unstable training and are generally hard to train well. How can you confidently ensure that this behaviour isn't due to the RL policy not being trained for long enough?

Three knowledgeable referees agree that the paper makes a valuable contribution to the MRI acceleration and reconstruction literature. They recognize the soundness of the proposed approach and its compelling experimental validation. I agree with the referees and recommend acceptance.

In today's clinical practice, magnetic resonance imaging (MRI) is routinely accelerated through subsampling of the associated Fourier domain. Currently, the construction of these subsampling strategies - known as experimental design - relies primarily on heuristics. We propose to learn experimental design strategies for accelerated MRI with policy gradient methods. Unexpectedly, our experiments show that a simple greedy approximation of the objective leads to solutions nearly on-par with the more general non-greedy approach. We offer a partial explanation for this phenomenon rooted in greater variance in the non-greedy objective's gradient estimates, and experimentally verify that this variance hampers non-greedy models in adapting their policies to individual MR images.

Test-time data augmentation---averaging the predictions of a machine learning model across multiple augmented samples of data---is a widely used technique that improves the predictive performance. While many advanced learnable data augmentation techniques have emerged in recent years, they are focused on the training phase. Such techniques are not necessarily optimal for test-time augmentation and can be outperformed by a policy consisting of simple crops and flips. The primary goal of this paper is to demonstrate that test-time augmentation policies can be successfully learned too. We~introduce \emph{greedy policy search} (GPS), a simple but high-performing method for learning a policy of test-time augmentation. We demonstrate that augmentation policies learned with GPS achieve superior predictive performance on image classification problems, provide better in-domain uncertainty estimation, and improve the robustness to domain shift.