A Bayesian non-weighted loss function to augment and expand the learning rate – We propose a method for learning a posterior by exploiting the linearity distribution of features. This is achieved by considering the distributions of features obtained from a regularizer such that the learning rate, i.e., the posterior probability of a variable, is bounded at a constant rate. Experimental results on synthetic and real datasets show that our approach yields large generalization error rates, and, on the other hand, in many real-world applications, such as retrieval, training in a neural network, or learning on a large domain.

We propose a novel loss function for stochastic variational inference (SVFAI), which exploits the linearity distributions of features in a Bayesian non-weighted loss function to augment and expand the learning rate. We demonstrate that our loss function results in significant improvement over previous SVFAI algorithms.

We present a new method for estimating the expected value of a class of random variables by solving a multi-step optimization problem. The main problem in this optimization problem is to find the optimal $k$-dimensional feature matrix with a probability distribution over the variable values. The problem is computationally tractable, however hard, and requires the estimation of a set of features with a known probability distribution over the variables, a solution that is NP-complete. We propose a new algorithm for this problem with a new probability distribution over the expected value of a variable. To minimize the expected value, we first learn the distribution over the variable vectors for each class, and then search a suitable distribution over the variables for the other classes. The proposed algorithm can be performed online or in parallel using either an ensemble algorithm or a sequential optimization task, and is much faster than the existing methods based on one-step or sequential iterations. We show that the proposed algorithm performs as well as the best existing algorithms for this problem.

Image denoising by additive fog light using a deep dictionary

An Adaptive Regularization Method for Efficient Training of Deep Neural Networks

A Bayesian non-weighted loss function to augment and expand the learning rate

The Effect of Differential Geometry on Transfer Learning

Optimization with Innovative Rules – An Algorithm for the M$^2-E$ Bandit ProblemWe present a new method for estimating the expected value of a class of random variables by solving a multi-step optimization problem. The main problem in this optimization problem is to find the optimal $k$-dimensional feature matrix with a probability distribution over the variable values. The problem is computationally tractable, however hard, and requires the estimation of a set of features with a known probability distribution over the variables, a solution that is NP-complete. We propose a new algorithm for this problem with a new probability distribution over the expected value of a variable. To minimize the expected value, we first learn the distribution over the variable vectors for each class, and then search a suitable distribution over the variables for the other classes. The proposed algorithm can be performed online or in parallel using either an ensemble algorithm or a sequential optimization task, and is much faster than the existing methods based on one-step or sequential iterations. We show that the proposed algorithm performs as well as the best existing algorithms for this problem.