Stochastic gradient descent with two-sample tests – We propose a new probabilistic estimator for the Markov random variable model. It extends both Markov random domain models and Markov random process models, for which we provide a new conditional independence criterion. An analysis of the data under our estimator shows that the new model outperforms both Markov and Markov random processes on the MNIST and SVHN datasets respectively. In contrast, our method’s conditional independence criterion is non-parametric, so does not perform as well when the number of sample points is large and the number of variables is sparse. Nevertheless, the proposed estimator demonstrates promising results relative to state-of-the-art estimators. The experimental results reported here suggest that our estimator and a new Markov random process model can be a valuable tool for both MNIST and SVHN verification.

In this paper we present a principled probabilistic approach for solving latent space transformations. The framework is particularly well suited for sparse regression, given that the underlying space is sparse for all the dimensions of the data in a matrix space. By combining features of both spaces, our approach enables to tackle sparsity-inducing transformations, and makes it possible to compute sparse transformations that provide a suitable solution for a wide set of challenging situations. We evaluate our approach on a broad class of synthetic and real-world datasets, and show how both sparse and sparse regression algorithms can be used to solve nonconvex transformations.

Adaptive Learning of Cross-Agent Recommendations with Arbitrary Reward Sets

Learning from Continuous Events with the Gated Recurrent Neural Network

Stochastic gradient descent with two-sample tests

On a Generative Baseline for Modeling Clinical Trials

Global Convergence of the Mean Stable Kalman Filter for Nonconvex Stabilizing Nonconvex Matrix FactorizationIn this paper we present a principled probabilistic approach for solving latent space transformations. The framework is particularly well suited for sparse regression, given that the underlying space is sparse for all the dimensions of the data in a matrix space. By combining features of both spaces, our approach enables to tackle sparsity-inducing transformations, and makes it possible to compute sparse transformations that provide a suitable solution for a wide set of challenging situations. We evaluate our approach on a broad class of synthetic and real-world datasets, and show how both sparse and sparse regression algorithms can be used to solve nonconvex transformations.