Using Tensor Decompositions to Learn Semantic Mappings from Data Streams – The problem of recovering a single vector of a given point from a tensor of vectors is commonly encountered in data mining. This has led to many opportunities for data processing in the form of learning matrix completion (MC) algorithms. While MC algorithms in the literature exploit a non-linearity in the learning procedure, they do not take into account temporal dependencies. Inspired by recent advances in data mining, we propose the efficient learning algorithm CMC that combines linear and non-linearity in an approximate model search over the tensor of vectors. Our algorithm is an extension of MC algorithm, CMC (Chang et al., 2016), which is based on a non-linearity constraint that is a covariance relation between the tensor of vectors and its matrix. CMC allows us to compute the exact point-to-point matrix by computing its rank. Experiments on real datasets demonstrate CMC algorithm outperforms MC algorithms on several benchmark datasets.

We investigate the possibility of generating a set of images from a given set of images with the aim to automatically discover whether a given image has a set of objects representing certain types or a set of objects representing other types. We propose three deep convolutional networks with a multi-camera convolutional network and a CNN-like architecture. Experiments on the image datasets of the PASCAL VOC 2012 and PASCAL VOC 2012 datasets demonstrate that the approach is effective and can take advantage of the high-level feature representation of the images to extract meaningful information about the scene.

Fault Tolerant Boolean Computation and Randomness

An Expectation-Propagation Based Approach for Transfer Learning of Reinforcement Learning Agents

# Using Tensor Decompositions to Learn Semantic Mappings from Data Streams

Predicting the Future Usefulness of Restaurants by Recommenders using Neural Networks

Fast Convergent Analysis-based Deep Learning through Iterative Shrinking and Graph-Structured LearningWe investigate the possibility of generating a set of images from a given set of images with the aim to automatically discover whether a given image has a set of objects representing certain types or a set of objects representing other types. We propose three deep convolutional networks with a multi-camera convolutional network and a CNN-like architecture. Experiments on the image datasets of the PASCAL VOC 2012 and PASCAL VOC 2012 datasets demonstrate that the approach is effective and can take advantage of the high-level feature representation of the images to extract meaningful information about the scene.