This paper presents an extension to the popular metalearning algorithm MAML, in which it is re-cast as inference in a graphical model. This framing allows samples to be drawn from a model posterior, enabling reasoning about uncertainty and capturing multiple modes of ambiguous data, while MAML can only make a single point estimate of model parameters at test time. This is shown in several experiments to better capture the characteristic of ambiguous, noisy data than MAML. Strengths: The paper makes a strong point that few shot learning is often too ambiguous to confine to a single-model metalearning paradigm. Especially with the high level of recent interest in topics such as safe learning, risk-aware learning, and active learning, this is a relevant area of work.

This paper presents a reparametrization method for inference in undirected graphical models. At the heart of the method is the observation that all non-unary potentials can be made symmetric, and once this has been done, the unary potentials can be changed to pairwise potentials to obtain a completely symmetric probability distribution, one where x and its complement have the exact same probability. This introduces one extra variable, making the inference problem harder. However, we can now "clamp" a different variable from the original distribution, and if we choose the right variable, approximate inference might perform better in the new model than in the original. Inference results in the reparametrized model easily translate to inference results in the original model.

Update: I downgrade my review to 5. The main concern is 1) Some more extensive simulations will make the results more convincing, as the numerical experiment is the only way to assess the performance of the proposed priors. It might take a major revision to reflect such comprehensive comparisons. With that being said, I believe the paper does contain interesting results that are novel and useful to the community. In particular, the theoretical results seem sound, and the paper is fairly readable. But I think there is also room for improvement.

The paper proposes a modification of a reinforcement learning (RL) framework, called Value-Aware Model Learning (VAML), that makes the associated optimization problem more tractable. VAML is a model-based approach that takes into account the value function while learning the model. In its original formulation, VAML poses the problem as a "min max" optimization in which one seeks a model considering the worst-case scenario over the space of representable value functions. This paper proposes to replace the problem above with a sequence of optimizations whose objective functions include the actual value-function approximations that arise in value iteration (that is, one replaces the "max" above with a sequence of concrete approximations). The paper presents a theoretical analysis of the proposed method, first providing finite sample guarantees for the model-based approximation, then providing a general error propagation analysis, and finally combining the two.

Strengths 1. Considering dynamic programming problems in continuous time such that the methodologies and tools of dynamical systems and stochastic di_x000b_eren- tial equations is interesting, and the authors do a good job of motivating the generalities of the problem context. The parameterizations considered of the value functions at the end of the day belong to discrete time, due to the need to discretize the SDEs and sample the state-action-reward triples. Given this discrete implementa- tion, and the fact that experimentally the authors run into the conven- tional di_x000e_culties of discrete time algorithms with continuous state-action function approximation, I am a little bewildered as to what the actual bene_x000c_t is of this problem formulation, especially since it requires a re- de_x000c_nition of the value function as one that is compatible with SDEs (eqn. That is, the intrinsic theoretical bene_x000c_ts of this perspective are not clear, especially since the main theorem is expressed in terms of RKHS only. However, these methods are fundamentally limited by their sample complexity bottleneck, i.e., the quadratic complexity in the sample size.

This paper proposed a model powered approach to conduct conditional independent tests for iid data. The basic idea is to use nearest neighbor bootstrap to generate samples which follow a distribution close to the f {CI} and a classifier is trained and tested to see if it is able to distinguish the observation distribution and the nearest neighbor bootstrapped distribution. If the classification performance is close to the random guess, one fails to reject the null hypothesis that data follows conditional independence otherwise one accept the null hypothesis. In general, the paper is trying to address an important problem and the paper is presented in a clear way. It seems that the whole method can be decoupled into two major components.

This paper designs a model for conditional density estimation. It resembles a VAE architecture, where both x and y are given as inputs to the encoder to produce a latent variable w. W and x are then fed to the decoder to produce p(y x). However, unlike in VAE, x and y are not single data points, but rather sets and the decoder part uses GPs to output p(y x). I found the clarity of the paper very low and I wish authors explained the model in Section 3.1. Figure 1 made me especially confused as I initially thought that the model receives a single datapoint (x,y) just like a VAE.

Effective Bayesian Modeling of Groups of Related Count Time Series.

This paper proposes a Bayesian model for adversarial learning problem. Empirical studies on Fashion-MINST and traffic sign recognition show that the proposed methods is slightly better than other adversarial learning baselines. Below I list my concerns about the paper: For modeling, 1. This paper ignore a highly relevant work'Bayesian GAN' [1]. The non-cooperative game between'data generator' and'learner' established in this paper is almost the same as the vanilla GAN.

Overview: This paper looks at nonparametric density estimation under certain classes of metrics, which the authors call "adversarial losses". To define adversarial losses, assume we have two probability distributions P and Q, and suppose X P and Y Q. Consider the supremum of E[f(X)] - E[f(Y)], when f ranges over a class F_d of functions. This is the "adversarial loss with respect to class F_d", or simply the "F_d-metric", and generalizes several metrics, for instance the L_1 metric. Now, assume P is an unkown distribution belonging to some known class F_G, and we have n i.i.d. How small can we make this distance?