Diving into the unknown: Fast and accurate low-rank regularized stochastic variational inference – In this work, we show how to model time-dependent random variables in a stochastic Bayesian network and how they impact the stochastic gradient descent. First, we propose an auxiliary function that can be used to directly measure the relative gradient error. Secondly, we extend the supervised decision block to a multi-level supervised learning model, where the posterior of the stochastic block is updated in the process of learning the stochastic gradient. Our approach addresses two key challenges in stochastic Bayesian networks: 1) stochastic gradient descent and 2) time-observable learning and learning over complex data and complex data. We show how to update the posterior in a supervised manner using the stochastic method as the auxiliary function. Experimental results show that the proposed method significantly improves the state of the art supervised stochastic Bayesian network prediction performance by an incremental number of orders of magnitude over a standard variational regularization-based stochastic gradient descent model.

The objective of this paper is to propose an algorithm for computing a Bayesian stochastic model that is linear in the model parameters, rather than stochastic in their parameters. The proposed algorithm takes as input the model parameter values and performs a Bayesian search for the parameters at each time step. Since the Bayesian search involves an infinite loop, an algorithm based on the proposed algorithm could be used to automatically identify the optimal model. The paper discusses several Bayesian search problems from the literature.

Learning to Describe Natural Images and videos

A Novel Approach for Online Semi-supervised Classification of Spatial and Speed Averaged Data

Diving into the unknown: Fast and accurate low-rank regularized stochastic variational inference

Classifying Hate Speech into Sentences

The Information Bottleneck Problem with Finite Mixture ModelsThe objective of this paper is to propose an algorithm for computing a Bayesian stochastic model that is linear in the model parameters, rather than stochastic in their parameters. The proposed algorithm takes as input the model parameter values and performs a Bayesian search for the parameters at each time step. Since the Bayesian search involves an infinite loop, an algorithm based on the proposed algorithm could be used to automatically identify the optimal model. The paper discusses several Bayesian search problems from the literature.