Thus, we now know which ad/query pairs are shoppy or not.

In our article, we introduce a Riemannian batch normalization (batch-norm) algorithm, which generalizes the one used in Euclidean nets.

MLPs, which we will call "coordinate-based" MLPs, take low-dimensional coordinates as inputs

Q2: Please summarize your review in 1-2 sentences The paper introduces a parallelizable split merge MCMC algorithm for the HDP, suitable for corpora as large as the NYTimes dataset. The inference method is compared against two other inference algorithms for the HDP, demonstrating better convergence properties.

Some commonly used objective loss function such as cross-entropy loss for classification and squared loss for regression can be regarded as special cases of the logarithmic loss function.

We sincerely thank Reviewer 1 for referring us to four relevant papers [1-4]. Paper [1] provides a very interesting relationship between Fisher divergence and Stein's operator, whereby Hyvarinen Paper [4] establishes a more general result than that of S. If the model class is well-specified then convergence to the data generating distribution is guaranteed. Fano's inequality also gives lower bounds of model selection/message decoding error (so larger'A kernelized Stein discrepancy for goodness-of-fit tests.' We appreciate Reviewer 2's comments and recommendations. We will do another proof reading and remove typos.

Neural Information Processing Systems http://nips.cc/