We thank all the reviewers for their constructive comments. Please see our responses below. Our approach is purely based on 2D convolutions. We thank the reviewers for pointing out some related (or missing) references. Timeception, SlowFast and TSM are concurrent with our work.

We will make this more clear in the paper.

We thank the reviewers for their careful reading, feedback and helpful comments and address specific concerns below. Tasks are best retained using the double-sided approach. In the revised paper, we will make it more explicit that analyses of sections 5.3-5.5 were We agree that saying "were allowed" in line 195 is R1,3 noted that our set of tasks were limited. We view this as an important first step, but agree that real-world applications (e.g. Fixed point structures were highly overlapping upon visual inspection in TDR subspaces.

First, our paper is not a statement that Graph-Q-SA T is a state of the art SA T heuristic.

We thank the reviewers for their feedback and their time. R2: "MAP can give more accurate predictions". Specifically, [11] shows overfitting to data yields better accuracy. "Bayesian Deep Learning and a Probabilistic Perspective of Generalization", section 9.1), which becomes more apparent For MF-VI, many units' variances converge to values close to prior variance, as shown in Figure 1 We will improve this writing in the paper. We also note that k-tied + our approach will yield higher ELBO than k-tied alone.

We thank all the reviewers for their constructive comments. Please see our responses below. Our approach is purely based on 2D convolutions. We thank the reviewers for pointing out some related (or missing) references. Timeception, SlowFast and TSM are concurrent with our work.

Post-rebuttal: The authors have promised to incorporate an exposition of the sampler in the revised paper, I believe that will make the paper a more self-contained read. I maintain my rating of strong accept (8). I think this paper makes very nice contributions to the fundamental question of estimating the MLE distribution given a bunch of observations. I think the key contributions can be broken up into two key parts: - A bunch of simple but elegant structural results for the MLE distribution in terms of'tent distributions' -- distributions such that its log-density is piecewise linear, and is supported over subdivisions of the convex hull of the datapoints. This allows them to write a convex program for optimizing over tent distributions.

Visualization enables effective data exploration by harnessing the higher bandwidth interactivity of the human visual cortex. But the space of possible visualizations is immense, such that general abstractions for creating (that is, finding) the right visualization are elusive, despite recent progress in systems like vega2 and Draco.1 The following paper provides a general abstraction, along with advanced interfaces, focusing on visualization search. If you have ever created a long sequence of visualizations looking for interesting patterns, you have manually performed a visualization search task. The visualization search problem is to find subsets of the data that, when suitably rendered, generate a visualization like a provided pattern specification. This task is intuitively difficult, requiring at least a model of visualization similarity, a representation of a massive search space, a strategy for navigating the search space, and appropriate interfaces through which users can express specifications.