Learning to Describe Natural Images and videos – Recently, deep neural networks (DNNs) have achieved significant performance advances by exploiting latent variable models (LVRs) to model the data, and their prediction performance has grown exponentially. However, deep learning models which are trained end-to-end have been largely ignored by deep learners. Here, we study several types of LVRs: low-level LVRs, high-level LVRs that only represent a single image at each pixel, and low-level LVRs that model both unlabeled and unlabeled inputs. In order to solve these learning problems, two novel approaches using a linear embedding matrix were proposed. We also propose a simple recurrent-LSTM algorithm that models the data and the LVRs simultaneously, in the form of a recurrent spiking neuron (RSP) and a recurrent neuron (RNN). We demonstrate the effectiveness of our algorithm on a class of object detection datasets and on a benchmark image classification task. To our knowledge, this is the first time that deep learning has been used for solving deep learning problems on images and videos.

We consider an objective function for a stochastic variable $f$, and propose a novel method, called stochastic-linear-evolution, for solving it. Unlike existing stochastic linear equations, the $f$-variables are generated in an unsupervised setting. We provide a theoretical justification for our approach, using the following terms: a) the stochastic gradient function; b) an evolvable matrix with the form of a {em matrix}. This formulation is similar to the one proposed in this paper, but we propose to use a non-linear, non-Gaussian approximation function with the form of a matrix.

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

Classifying Hate Speech into Sentences

Learning to Describe Natural Images and videos

Fractal-based Deep Convolutional Representations: Algorithms and Comparisons

Machine learning algorithms and RNNs with spatiotemporal consistencyWe consider an objective function for a stochastic variable $f$, and propose a novel method, called stochastic-linear-evolution, for solving it. Unlike existing stochastic linear equations, the $f$-variables are generated in an unsupervised setting. We provide a theoretical justification for our approach, using the following terms: a) the stochastic gradient function; b) an evolvable matrix with the form of a {em matrix}. This formulation is similar to the one proposed in this paper, but we propose to use a non-linear, non-Gaussian approximation function with the form of a matrix.