I like the fact that the method is very simple to understand and implement (see my summary), and does not require any major changes to the base SG-MCMC algorithm. Also, this seems very general and applies to a large class of SG-MCMC algorithms, and is therefore potentially very impactful to the Stochastic Gradient MCMC community. Novelty: Although Richardson-Romberg extrapolation is well known in numerical analysis, it is not widely known in the machine learning / stochastic gradient MCMC community. Clarity: The paper is well written and the presentation is clear. Comments/questions: - Can this technique be directly applied to all SG-MCMC methods?

Summary This paper proposes a control variate (CV) for the discrete distribution's REINFORCE gradient estimator (RGE). The CV is based on the Concrete distribution (CD), a continuous relaxation of the discrete distribution that admits only biased Monte Carlo (MC) estimates of the discrete distribution's gradient. Yet, using the CD as a CV results in an *unbiased* estimator for a discrete random variable's (rv) path gradient as well as lower variance than the RGE (as expected). REBAR is derived by exploiting the REINFORCE estimator for the CD and by observing that given a discrete draw, the CD's continuous parameter (z, here) can be marginalized out. REBAR has some nice connections to other estimators for discrete rv gradients, including MuProp.

The authors present a variational inference algorithm for continuous time Markov jump processes. Following previous work, they use "uniformization" to produce a discrete time skeleton at which they infer the latent states. Unlike previous work, however, the authors propose to learn this skeleton (a point estimate, via random search) and to integrate out, or collapse, the transition matrix during latent state inference. They compare their algorithm to existing MCMC schemes, which also use uniformization, but which do not collapse out the transition matrix. While this work is well motivated, I found it difficult to tease out which elements of the inference algorithm led to the observed improvement.

Title: Differentially private Bayesian learning on distributed data Comments: - This paper develops a method for differential privacy (DP) Bayesian learning in a distributed setting, where data is split up over multiple clients. This differs from the traditional DP Bayesian learning setting, in which a single party has access to the full dataset. The main issue here is that performing DP methods separately on each client would yield too much noise; the goal is then to find a way to add an appropriate amount of noise, without compromising privacy, in this setting. To solve this, the authors introduce a method that combines existing DP Bayesian learning methods with a secure multi-party communication method called the DCA algorithm. Theoretically, this paper shows that the method satisfies differential privacy.

This paper presents a sampling method that combines Hamiltonian Monte Carlo (HMC), mini-batches, tempering, and thermostats, to more efficiently explore multimodal target distributions. It is demonstrated on a number of substantial neural network problems using real data sets. This is an interesting method, and the empirical results are quite substantial. Figure 2 does a nice job of demonstrating how the omission of any of the ingredients (e.g. the tempering, or the thermostat) is detrimental to the overall result, which is a nice illustration of how the combination works together well. This is followed by some substantial image classification examples.

This submission proposes a new prior on the topic-word distribution in latent topic models. This model defines a multi-layer feedforward graph, where each layer contains a set of valid multinomial distributions over the vocabulary, and weighted combinations of each layer's "topics" are used as the Dirichlet prior for the "topics" of the next layer. The key purported benefits are sharing of statistical strengh, inference of a hierarchy of interpretable "abstract" topics, and modularity that allows composition with other topic model variants that modify the document-topic distributions. The authors present an efficient fully collapsed Gibbs sampler inference scheme - I did not thoroughly check the derivation but it seems plausible. Although: what is the computational complexity (and relative "wall clock" cost) of the given inference scheme?

The paper presents a policy gradient algorithm for a multiagent cooperative problem, modeled in a formalism (CDEC-POMDP) whose dynamics, like congestion games, depend on groups of agents rather than individuals. This paper follows the theme of several similar advances in theis field of complex multiagent planning, using factored models to propose practical/tractable approximations. The novelty here is the use of parameterized policies and training algorithms inspired by reinforcement learning (policy gradients). The work is well-motivated, relevant, and particularly well-presented. The theoretical results are new and important.

The authors study the problem of structure learning for Bayesian networks. The conventional methods for this task include the constraint-based methods or the score-based methods which involve optimizing a discrete score function over the set of DAGs with a combinatorial constraint. Unlike the existing approaches, the authors propose formulating the problem as a continuous optimization problem over real matrices, which performs a global search, and can be solved using standard numerical algorithms. The main idea in this work is using a smooth function for expressing an equality constraint to force acyclicity on the estimated structure. The paper is very well written and enjoyable to read.

Summary: Meta-learning is motivated by the promise of being able to transfer knowledge from previous learning experiences to new task settings, such that a new task can be learned more effectively from few observations. Yet, updating highly parametric models with little amounts of data can easily lead to overfitting. A promising avenue towards overcoming this challenge is a Bayesian treatment of meta-learning. This work, builds on top of recent work that provides a Bayesian interpretation of MAML (model-agnostic-meta-learning). This contribution is a direct extension of (Grant et al 2018) - where the task-train posterior was approximated via a Gaussian distribution. Applying SVGD instead allows for a more flexible and (potentially) more accurate approximation of a highly complex posterior.

The authors propose a marked process latent variable model that leverages Gaussian processes for continuous latent sates, a generalized linear model for discrete latent states and an inference network for efficient inference. Results on real data suggest that the proposed approach is interpretable and outperforms standard baselines on held out data. I really enjoyed reading the paper, it is very well written. That being said, the notation is sloppy and lack of details makes it difficult to appreciate the contributions of the paper. The authors seemed to have assumed that the reader is very familiar with point processes, Gaussian processes, variational inference and deep learning.