Additional Feedback: - At equation (2) M should be defined.

Weaknesses: There are two weaknesses with the paper: 1. I am not sure if the theoretical claims are correct. They seem very strong as only in propositional case, we have FPT for constant treewidth, so the claims need to talk about what makes problem so hard when you add in continuous variable. I looked at the proof and there are several things that trouble me: There is change of representation; subset sum is #P-complete when we are concerned with binary representation otherwise we have pseudo-polynomial time algorithms by dynamic programming. The proofs seem to work on integer representation not binary representation as the treewidth for binary representation is still n.

Two reviewers rated the paper very highly (8, 9). R3 initially gave a very low rating and expressed concerns about correctness of the intractability result, especially related to the representation of numbers in the reduction from subset sum. This point received significant discussion. The meta-reviewer read the proof and felt it was very clear and agrees with the authors. There is one minor ambiguity that can be resolved: the authors don't specify the representation of the constants that appear in the constraints in the WMI problem instance.

Overall the paper is an average paper but clearly written. This paper proposes an improvement of Charikar's approach to achieve sublinear kernel density estimation with linear space and linear time preprocessing. Experimental results focus mainly on Laplacian (L1 variant in the main submission and L2 variant added in supplement). The key observation for achieving linear space is to modify the previous HBE approach so that each hash table stores each point in the dataset with constant probability - in this way, the superlinear storage cost is overcome. However, my main complaint is in the experimental results.

After a careful discussion among the reviewers, there is a clear consensus that the paper provides a solid theoretical contributions for kernel density estimation in high dimensions. Hence, I am happy to recommend to accept this paper for publication at NeurIPS2019. Nevertheless, one concern that came up during the discussion is the lack of experimental comparison and clarity in how it was performed. I urge the authors to incorporate reviewers' comment on this weakness of the paper into the camera-ready version.

This paper deals with an important problem in theoretical reinforcement learning (RL), that is, finite-time analysis of on-policy RL algorithms such as SARSA. If the analysis techniques, as well as proofs, were correct and concrete, this work may have a broad impact on analyzing related stochastic approximation/RL algorithms. Although important and interesting, the present submission contains several major concerns, that have limited the contributions and even brought into question the practical usefulness of the reported theoretical results. These concerns are listed as follows. To facilitate analysis, a number of the assumptions adopted in this work are strong and impractical.

Because the initial reviews were mixed, I obtained an additional review from an expert in the area of this paper. This 4th review came back clearly positive, but in the mean time one of the positive reviewers changed to negative (and later one of the negatives turned to positive). Then we had a lot of discussion, but the reviewers never did agree on how best to view this paper. In fact, they seemed to talk past each other, and in the end we had two positive and two negative reviews. As the area chair, reading the reviews and listening to the discussion, I found the 4th, very-positive review to be the most compelling.

The paper considers how to create a flexible method for modelling Hawkes-like processes with flexibility in the triggering kernel, using Gaussian processes working on a different input than is usually attempted. Section 1-3 is well written, very clear and gives a good motivation and description of how to get a GP regressive point process. The idea in the preliminary work that the likelihood of a Hawkes process factorises across different types is surprising as the triggering kernel in the CIF appears to connect \lambda_{u_{i}}(t) and \lambda_{u_{j}}(t), since u_{i} and u_{j} appear to be part of a single triggering kernel, and hence are connected. Can the authors clarify, is there some conditional independence I am not spotting here? It doesn't appear to be of significance however as the rest of the paper only focusses on a single event type in the end.

Active Learning (AL) is a sequential learning approach aiming at selecting the most informative data for model training. In many systems, safety constraints appear during data evaluation, requiring the development of safe AL methods. Key challenges of AL are the repeated model training and acquisition optimization required for data selection, which become particularly restrictive under safety constraints. This repeated effort often creates a bottleneck, especially in physical systems requiring real-time decision-making. In this paper, we propose a novel amortized safe AL framework. By leveraging a pretrained neural network policy, our method eliminates the need for repeated model training and acquisition optimization, achieving substantial speed improvements while maintaining competitive learning outcomes and safety awareness. The policy is trained entirely on synthetic data utilizing a novel safe AL objective. The resulting policy is highly versatile and adapts to a wide range of systems, as we demonstrate in our experiments. Furthermore, our framework is modular and we empirically show that we also achieve superior performance for unconstrained time-sensitive AL tasks if we omit the safety requirement.

Federated learning collaboratively trains a neural network on a global server, where each local client receives the current global model weights and sends back parameter updates (gradients) based on its local private data. The process of sending these model updates may leak client's private data information. Existing gradient inversion attacks can exploit this vulnerability to recover private training instances from a client's gradient vectors. Recently, researchers have proposed advanced gradient inversion techniques that existing defenses struggle to handle effectively. In this work, we present a novel defense tailored for large neural network models. Our defense capitalizes on the high dimensionality of the model parameters to perturb gradients within a subspace orthogonal to the original gradient. By leveraging cold posteriors over orthogonal subspaces, our defense implements a refined gradient update mechanism. This enables the selection of an optimal gradient that not only safeguards against gradient inversion attacks but also maintains model utility. We conduct comprehensive experiments across three different datasets and evaluate our defense against various state-of-the-art attacks and defenses. Code is available at https://censor-gradient.github.io.