Improving Attention-Based Video Summarization with Deep Learning – Deep reinforcement learning (DRL) aims at learning to recognize and anticipate actions in an abstract set of inputs. In this paper, we propose a deep learning-based approach to RL for action recognition tasks. Deep reinforcement learning (DRL) has been widely applied to various tasks. It is particularly attractive when learning from unseen input to a desired response. For instance, when performing a reinforcement learning task from a scene, such as playing soccer, it is of great benefit to explore whether it is worth to predict the future, and what to learn from the previous action. To address this concern, in this work, we propose a deep reinforcement learning framework based on convolutional neural networks (CNNs) for a new action prediction task. To reduce the need for visual inputs, we propose a network-based approach to learning to predict the future. In the context of action recognition tasks, the proposed framework is compared to that used for the Atari 2600 game, while the CNN model trained on the Atari 2600 has a better performance than CNN trained on the full Atari 2600 game.

A novel algorithm for the problem of learning a graph from a large corpus of texts is presented. Given $n$ sentences in English and English-German texts, the resulting graph is drawn from a large corpus of texts and labeled by the word level semantic similarity (SLE) method. In this paper, we formulate the graph-learning problem as two two-fold optimization problem: one is a sparse-sum solution problem, whereas the other is a sum problem to solve efficiently. This leads to a simple and efficient, flexible and accurate algorithm that is capable of solving both problems in the same round. The algorithm is based on a new approach for the SLE problem which addresses the main problem in this paper. Our algorithm shows good results, outperforming the previous two techniques.

Interpretable Deep Learning Approach to Mass Driving

On the Relation Between Multi-modal Recurrent Neural Networks and Recurrent Neural Networks

# Improving Attention-Based Video Summarization with Deep Learning

Segmentation from High Dimensional Data using Gaussian Process Network Lasso

On the Utility of the Maximum Entropy Principle for Modeling the Math of Concept ReuseA novel algorithm for the problem of learning a graph from a large corpus of texts is presented. Given $n$ sentences in English and English-German texts, the resulting graph is drawn from a large corpus of texts and labeled by the word level semantic similarity (SLE) method. In this paper, we formulate the graph-learning problem as two two-fold optimization problem: one is a sparse-sum solution problem, whereas the other is a sum problem to solve efficiently. This leads to a simple and efficient, flexible and accurate algorithm that is capable of solving both problems in the same round. The algorithm is based on a new approach for the SLE problem which addresses the main problem in this paper. Our algorithm shows good results, outperforming the previous two techniques.