Convex Relaxation Learning – We present a method for solving a learning-based, Bayesian network in which the inputs are random vectors of the number of possible outcomes. This is a special case of linear programming due to its nonconvex form. This formulation is not optimal in the context of learning with linear programs, where the inputs should consist of both samples and vectors. To improve the model, the parameters of the network must belong to a different set (the number of outcomes). However, this is a common problem to solve, and a general rule of thumb is that the learning algorithm must only satisfy some constraints. In this paper, we solve the problem in the Bayesian framework, by means of a stochastic approximation. In this way, the learning algorithm can be reduced to the classical stochastic approximation, which requires a priori knowledge of the probability distributions of the variables. Since the nonconvex formulation is not optimal, we prove that the stochastic approximation is optimal for a learning-based, Bayesian network, where the inputs are random vectors of the number of possible outcomes.

Sparse-time classification (STR) has emerged as a promising tool for automatic vehicle identification. The main drawback of STR is its lack of training data and the difficulty of handling noisy data. In this work we present an innovative approach to the problem using Convolutional Neural Networks. In our model, we first use unsupervised learning as feature representation for image classification: the Convolutional Neural Network (CNN) is trained with an unlabeled image. The CNN learns a binary metric feature embedding representation of its output vectors (e.g., the k-dimensional). Following this representation, the CNN can model the training data by selecting a high-quality subset of the training data. Our method learns the representations and, by using the learned representations, can be used with the standard segmentation and classification algorithms in order to learn the feature representation for the given dataset. We evaluate our method on the challenging TIDA dataset and compare it to the state-of-the-arts.

Story highlights An analysis of human activity from short videos

Learning from the Fallen: Deep Cross Domain Embedding

# Convex Relaxation Learning

Neural-based Word Sense Disambiguation with Knowledge-base Fusion

Machine Learning Methods for Multi-Step Traffic AcquisitionSparse-time classification (STR) has emerged as a promising tool for automatic vehicle identification. The main drawback of STR is its lack of training data and the difficulty of handling noisy data. In this work we present an innovative approach to the problem using Convolutional Neural Networks. In our model, we first use unsupervised learning as feature representation for image classification: the Convolutional Neural Network (CNN) is trained with an unlabeled image. The CNN learns a binary metric feature embedding representation of its output vectors (e.g., the k-dimensional). Following this representation, the CNN can model the training data by selecting a high-quality subset of the training data. Our method learns the representations and, by using the learned representations, can be used with the standard segmentation and classification algorithms in order to learn the feature representation for the given dataset. We evaluate our method on the challenging TIDA dataset and compare it to the state-of-the-arts.