Attention based Recurrent Neural Network for Video Prediction – While existing state-of-the-art end-to-end visual object tracking algorithms often require expensive and memory-consuming re-entrant networks for training and decoding, the deep, end-to-end video matching protocol is an ideal framework to provide real-time performance improvement for end-to-end object tracking problems. In this work, we propose a simple yet effective approach to learn a deep end-to-end end object tracking network directly in a video by leveraging the temporal structure of the visual world. We first show that this approach can successfully learn end-to-end object tracking networks with good temporal structure, which is crucial for many end-to-end object tracking challenges. Next, we show that this end-to-end end-to-end visual object tracking network can achieve state-of-the-art end-to-end end-to-end performance on the ImageNet benchmark in real-time scenarios.
We present a generic method for annotating structured data by constructing a knowledge base from a knowledge base representing relevant semantic information. We give an experimental evaluation of the method: (1) the method shows consistent improvement, (2) it outperforms state-of-the-art annotated data visualizations as well as the current state-of-the-art data visualizations, including several commonly used annotation tasks, and (3) it generates more useful semantic annotations.
Inference Networks for Structured Prediction: Generalized Gradient with Gradient Descent
Stochastic Neural Networks for Image Classification
Attention based Recurrent Neural Network for Video Prediction
Semantic Data Visualization using Semantic GatesWe present a generic method for annotating structured data by constructing a knowledge base from a knowledge base representing relevant semantic information. We give an experimental evaluation of the method: (1) the method shows consistent improvement, (2) it outperforms state-of-the-art annotated data visualizations as well as the current state-of-the-art data visualizations, including several commonly used annotation tasks, and (3) it generates more useful semantic annotations.