Video Frame Interpolation via Joint Determinantal and Dose Coding – We present a method for learning two-dimensional conditional probability distribution functions on a continuous set of frames. This is in contrast to several existing approaches, including the recently proposed approach for conditional probability estimation in the sense of conditional random field. The aim of the new approach is to learn a conditional probability distribution on the observed frames in order to obtain a more efficient conditional estimator. We illustrate that this approach improves the performance of conditional regression in a simulated application, where we show that the number of frames on test set are much smaller than those on typical real world images.
In this paper, we propose a novel deep neural network-based framework for decision making problems that involve multiple states in the state space. As a result, this framework offers new ways to interact with the state space through a simple feature selection procedure and a deep neural network learning framework. The framework is built on a deep neural network architecture and a recurrent neural system, a framework that can be trained from a single training example. To further facilitate the learning process of the framework, the framework is used as a training network on the state space. Our learning model allows us to design a new framework for solving multi-state planning problems, where multiple states are coupled into a single state by a single state. We demonstrate that our framework provides a method of solving problems that are asymptotically simple, but have a great complexity. The framework is able to handle a large variety of multi-state planning problems.
Efficient Stochastic Dual Coordinate Ascent
Scalable Bayesian Matrix Completion with Stochastic Optimization and Coordinate Updates
Video Frame Interpolation via Joint Determinantal and Dose Coding
Deep neural network training with hidden panels for nonlinear adaptive filtering
A Neural Approach to Reinforcement Learning and Control of Scheduling ProblemsIn this paper, we propose a novel deep neural network-based framework for decision making problems that involve multiple states in the state space. As a result, this framework offers new ways to interact with the state space through a simple feature selection procedure and a deep neural network learning framework. The framework is built on a deep neural network architecture and a recurrent neural system, a framework that can be trained from a single training example. To further facilitate the learning process of the framework, the framework is used as a training network on the state space. Our learning model allows us to design a new framework for solving multi-state planning problems, where multiple states are coupled into a single state by a single state. We demonstrate that our framework provides a method of solving problems that are asymptotically simple, but have a great complexity. The framework is able to handle a large variety of multi-state planning problems.