To Reviewers, we will make all suggested minor corrections in the final version and address main concerns below. This provides new perspectives to acceleration. In terms of experiments, SVR-ADA is compared with SOT A finite sum solvers. If we use one-norm, then it can only represent the general convex setting. In the final version, we will rewrite the abstract to make it more clear.

We thank all the reviewers for their insightful and constructive comments, and will revise the paper accordingly. We designed our model to match objects based on general principles (e.g., We stress that ADEPT's training was not specific to the test dataset: there were no We will release the dataset along with all code, human data, and model evaluations upon publication. We chose to model them separately to avoid producing a constant surprise signal. Observing the unexpected enhances infants' learning and exploration. Over-representation of extreme events in decision making reflects rational use of cognitive resources.

We thank the reviewers for their constructive comments. The two terms on the RHS of Eq. (13) are monotone increasing functions, and Using our Lemma 5.1's proof, Lemma 5.8 and Theorem 2's proof in Srinivas et al [19], the regret bound GP-UCB is chosen as it has ability to analyze convergence, which is very important in the unknown search space setting. EI convergence can be shown only in noiseless setting, PI/ES/PES do not have convergence proof yet). Thank you for the compliment though. We have conducted more experiments with the 10-dimensional functions Ackley10 and Levy10.

We thank all reviewers for the constructive comments. Examples include Xu et al., (2016, arXiv:1602.04511), Therefore, many existing methods adopt such a procedure. RMSE cannot distinguish which prediction is better, while the log-likelihood for the best prediction is the largest. RNN is not suitable for the short sequence data we are targeting, so we did not include it.

We thank the reviewers for their insightful and constructive comments. Reviewer #1 and shared comments. For latent FML we maximize the mutual information (Sec 3.3). We thank the reviewer for this very comprehensive review, which we really appreciate. We agree this point needs to be reinforced.

We thank the reviewers for their detailed and constructive comments, especially during these unprecedented times. Our algorithm isn't designed to compete (or However, in our new experiment in Fig. B we achieve close to D-SGD We will add to the paper an experiment with 4 different models. Reference data can be synthetic and then it is easy to obtain (as in co-regularization, see R1's comment). We now explain that in detail. The graphs in this work were randomly drawn for a given maximum number of degrees per node.

We would like to thank our reviewers for their constructive comments. R1: Not first to look at full meshes. Accordingly, we will limit our claim to deep learning approaches for human pose reconstruction. We will release those, along with pre-trained models. Why does fig 4 show stick figures?

An increase in the zero missingness probability of 1% leads to an approximately 0.28% increase in the final estimate of