First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper addresses the issue of object detection, in particular the challenge of obtaining bounding boxes on a scale similar to which category labels exist for object categorization. The authors side-step this challenge by proposing to adapt object classifiers for the detection task. Their algorithm is fairly simple and straightforward, which is not a bad thing in itself. Their experimental protocol uses 100 categories for training (with both category labels and bounding boxes), and tests on 100 left-out categories.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The paper studies the statistical consistency of plug in classifiers under non decomposable loss functions such as the F statistic which is a popular performance measure in machine learning. The problem studied in this paper is complex because non decomposable measures cannot, by definition, be expressed as an empirical expectation. Therefore, usual concentration inequalities are not applicable in this scenario. The authors present a general analysis for measures that can be expressed as a continuous function of the true positive rate and the true negative rate as well as the class probability.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper studies the worst case hardness of estimating the parameters of a binary pairwise undirected graphical model. By considering the specific case of the hard-core / independent set model and relying on the known result that approximating the partition function for this problem given the parameters (even for the unweighted case of all theta_i being 0) is hard, the authors show that the other direction -- namely, approximating the parameters given the node marginals -- is hard, in the sense that it does not admit an FPRAS. This is a strong theoretical paper, addressing a computational complexity problem that occurs very often in theory and practice, has been conjectured to be hard, but hadn't yet been shown formally to be hard (to the best of my and the authors' knowledge). The work is unlikely to have much impact in practice.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper proposes a denoising algorithm based on non-local image statistics and patch repetition by combining the advantages of NL-means and Exponentially Weighted Aggregation (EWA). The computation of the aggregated estimator is done using MCMC and results are comparable to state-of-the-art algorithms. Pluses: 1) the method seems simple and straightforward to implement. Minus: In terms of explaining *why* the method works, the text leaves something to be desired. E.g., in the second paragraph of section 7 (The proposed implementation proceeds in two identical iterations.)

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The paper presents a method for learning multiple tasks in parallel where at each round a sample is given per each task, but only a single task can have its sample annotated. The authors formulate their method using a trade-off between exploitation and exploration. Th1 provides an upper bound on the expected cumulative number of mistakes. The algorithm is compared to 2 different approaches for choosing the single sample/task to be annotated. It is well written and provides good theoretical as well as experimental results.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper's core finding is that combining an identity classification task as well as metric-learning-style verification task helps to learn better features for face classification/verification. The verification task here tries to decrease feature-space distance between instances of the same identity, and increase distance between those of different identities. This improvement is embedded in a state-of-the-art system for face verification, which uses convnets trained on many (400) different views to generate features, distilled into a small set of 25 using feature selection. Very good results are obtained and experiments performed using LFW as a test set. Overall, these are very good results obtained using a somewhat complex pipeline, and a good investigation into the contribution of each task in the loss for feature learning.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper proposes to learn bilingual word vector representations through an autoencoder. The novelty of this approach is to not rely on word-level alignments. It only requires aligned sentences. An autoencoder model is used to reconstruct the bag-of-words representation of aligned sentences, within and between languages.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The authors provide finite sample bounds on the excess risk of these classifiers. When taken to the limit these bounds reproduce the known consistency results for this class. However, they are superior in two ways: 1. They apply in the finite case 2. They apply to a broader set of metric spaces The presentation is very clear and the intuition is well described.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. The submission describes a convex deep learning formulation that leverages a number of key ideas. First, a training objective is proposed that explicitly includes the outputs of hidden layers as variables to be inferred via optimization. These are linked to linear responses via a loss function, and the net objective is the sum of these loss functions across the layers, plus some regularization terms. Next, a number of changes of variables are performed in order to reparameterize the objective into a convex form, heavily leveraging the representer theorem and the idea of value regularization. We are left with a convex objective in terms of three different matrices (per layer) to optimize. In particular, one of these matrices is a nonparametric'normalized output kernel' matrix, which takes the place of optimizing over the hidden layer outputs directly; however, this leads to a transductive method where we must simultaneously solve the optimization for training and test inputs.

"NIPS Neural Information Processing Systems 8-11th December 2014, Montreal, Canada",,, "Paper ID:","1407" "Title:","On Communication Cost of Distributed Statistical Estimation and Dimensionality" Current Reviews First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper investigates the communication cost of distributed estimation for d-dimensional spherical Gaussian distribution with unknown mean and unitary covariance, where the joint distribution is assumed to be a product distribution of each coordinate. The authors generalize previous works on the one-dimensional case in [4] by proposing upper and lower bounds for d-dimensional data on two communication schemes, interactive and simultaneous communication settings, for achieving minimax squared loss. The results establish the tradeoffs between dimensionality and communication cost for distributed estimation. In addition, improved bounds are derived when the unknown mean is s-sparse.