Neural Architectures of Visual Attention – We present an approach for the automatic retrieval of image features from a large-scale handwritten hand-annotated dataset. We present an algorithm, named ImageFRIE, based on Image-to-Image Encoding, which uses the deep convolutional neural network (CNN) to encode large-scale images into a small-scale one in terms of semantic features. In particular, we encode these features into a short segment that is fed to the deep CNN. The segment is learned and deployed to train a deep CNN, and to generate the labeled images of the label with which they align. We develop a novel neural network and demonstrate its ability to extract semantic information from hand-annotated images, and perform object recognition tasks in a single system (ImageFRIE). Our method is the first to use the CNN for this task. We demonstrate how the CNN can be used when extracting object features from hand-annotated hand-annotated data.

The recent success of deep learning in real-world problem solving was demonstrated for the sake of improving state-of-the-art model design and training for real world data. In this paper we propose a novel method to learn a representation for a probabilistic model to efficiently deal with uncertainty. Our results demonstrate that our model can learn representations of the probabilistic model, and our algorithm outperforms state-of-the-art models and models from the literature.

Automated prediction is a powerful tool to detect and to monitor the behavior behind a human activity in the wild. The goal of this work is to generate a predictive model of a player’s actions in a given scene in such a way that an effective prediction can be made. In this work we present a novel deep neural network model to learn the player’s behavior with a simple reinforcement learning framework. We demonstrate the importance of learning using our model by means of a realistic game.

Adversarial Methods for Robust Datalog RBF

An iterative k-means method for minimizing the number of bound estimates

Neural Architectures of Visual Attention

Nonconvex learning of a Bayesian network over partially observable domains

FlawBoost: Learning to Compare Models via Boosting on DistributionsThe recent success of deep learning in real-world problem solving was demonstrated for the sake of improving state-of-the-art model design and training for real world data. In this paper we propose a novel method to learn a representation for a probabilistic model to efficiently deal with uncertainty. Our results demonstrate that our model can learn representations of the probabilistic model, and our algorithm outperforms state-of-the-art models and models from the literature.

Automated prediction is a powerful tool to detect and to monitor the behavior behind a human activity in the wild. The goal of this work is to generate a predictive model of a player’s actions in a given scene in such a way that an effective prediction can be made. In this work we present a novel deep neural network model to learn the player’s behavior with a simple reinforcement learning framework. We demonstrate the importance of learning using our model by means of a realistic game.