P-Gauss Divergence Theory – Deep neural networks are highly capable of modeling information in a structured setting. However, the lack of suitable models to represent these forms of information does not explain their impressive performance. In this paper, we propose a new model that embeds the structured information in a fully connected Bayesian network structure. Specifically, we employ a Bayesian network structure to represent structured information. The model has been evaluated on various datasets, and it predicts the optimal model, i.e., the model with structured information, over the whole dataset. Our experimental results highlight the importance of learning these structures: We obtain consistent results for the optimal model and outperform all existing frameworks on both simulated and real datasets.

A new framework for sparsely-supervised learning (SSL) is proposed. The SSL framework is characterized by the use of multi-class sparsity for sparse learning. This sparsity framework is based on the use of a simple-to-sparse optimization procedure. The first-order optimization algorithm is used to estimate the parameters of the supervised learning model using the weighted sum of the observations. The second-order and the greedy algorithms are used to reduce the model size. The framework is developed to learn a low-rank sparsely-supervised algorithm by means of the greedy algorithm. Experimental evaluation shows that the framework is robust to the size and complexity of the sparsity and that the cost of the SG algorithm is reduced from $4^{-2}$ to $4^{-1}$.

Boosted-Autoregressive Models for Dynamic Event Knowledge Extraction

On a Generative Baseline for Modeling Clinical Trials

P-Gauss Divergence Theory

Training Discriminative Deep Neural Networks with Sparsity-Induced Penalty

On the Geometry of a Simple and Efficient Algorithm for Nonmyopic Sparse Recovery of Subgraph FeaturesA new framework for sparsely-supervised learning (SSL) is proposed. The SSL framework is characterized by the use of multi-class sparsity for sparse learning. This sparsity framework is based on the use of a simple-to-sparse optimization procedure. The first-order optimization algorithm is used to estimate the parameters of the supervised learning model using the weighted sum of the observations. The second-order and the greedy algorithms are used to reduce the model size. The framework is developed to learn a low-rank sparsely-supervised algorithm by means of the greedy algorithm. Experimental evaluation shows that the framework is robust to the size and complexity of the sparsity and that the cost of the SG algorithm is reduced from $4^{-2}$ to $4^{-1}$.