First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The paper proposes a new approach for Mahalanobis metric learning for k-nearest neighbor (kNN) classification. The main difference from the existing work is in the way how k nearest neighbors are found. Instead of simply looking for the k nearest neighbors, the authors are searching for the closest k examples that also guarantee correct classification (the authors refer to it as the gerrymandering). They propose a greedy algorithm to find such a neighborhood.

We thank the reviewers for their time and effort reviewing our paper. We are pleased that you found our work to "solve At the latest, source code will be made publicly available upon publication. When we write "greedy," we are referring to the sequential setting When we say "naive batch" we We will add a thorough discussion on this topic to a future version of the paper. We will update the section accordingly. The main issue is how to do this efficiently for complex, non-linear models.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This is an interesting and well written paper using variational approximations to deal with non-Gaussian likelihoods. The ideas in the paper are very well explained and the connections to the existing literature are good. The connections to the EKF and UKF are interesting. In section 1 and in section 3, it is claimed that the proposed methodology would be applicable to problems with non-factorizing likelihoods, but no further details are provided about this, and no experiments deal with this case.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper proposes an incremental but very sensible and practical modification to'curriculum learning'. Given a partition of the training examples into classes, they propose an additional regularising term (and an additional parameter) to ensure that the'easy' examples selected during learning are spread across the classes, and not from one class. The partition into classes can come from a clustering algorithm, or from a priori knowledge. The idea is straightforward and sensible, and the authors propose an algorithm that looks efficient and correct.

We will correct this inconsistency of terminology.

GPipe (with GD/AGD) may be seen as a special case of PPRS, with γ,K = 0 ( i.e., no randomized smoothing).