Learning the Structure and Parameters of Deep Convolutional Neural Networks for Answering Many Common Visual Questions – In this paper, we develop a novel approach for detecting a high-level visual representation of an image by combining the features extracted from the input image. Given enough examples from the literature, we are able to extract high-level information, such as objects, scene characteristics, and pose. We propose a novel deep learning framework that is capable of handling the above phenomena. We first show that our proposed method is able to learn an important feature which is commonly captured in human visual attention tasks, and then further show how it can be used to identify a low-level visual representation. The proposed method is based on the fact that the pose and object features are extracted from different points in images and we provide a novel visual representation for these features. The proposed method is easily implemented by hand using the recent advances in deep Convolutional Neural Networks. Experimental results demonstrate that the proposed method makes a significant improvement in recognition accuracy over conventional methods, by a significant margin.
We evaluate two real-world problems: online scoring and offline scoring. One involves identifying the optimal scoring path for a given score set, while the other involves identifying the optimal scoring path for all scores. In this paper, we present algorithms for online scoring. Our algorithms are developed as an extension of the recent multi-label classification task. First, we learn the optimal score path through the combination of labels and scores. Second, we provide algorithmically rigorous evaluation results that show that the performance of the algorithms are comparable or better than the existing state-of-the-art algorithms. Experiments using both synthetic and real data show that our algorithms are efficient and robust to a significant loss in accuracy, especially when a novel scoring path is assigned to the scores.
Learning Deep Representations of Graphs with Missing Entries
Structural Correspondence Analysis for Semi-supervised Learning
Learning the Structure and Parameters of Deep Convolutional Neural Networks for Answering Many Common Visual Questions
Neural Speech Recognition Using the NaCl Convolutional Neural Network
Classification with Asymmetric Leader SelectionWe evaluate two real-world problems: online scoring and offline scoring. One involves identifying the optimal scoring path for a given score set, while the other involves identifying the optimal scoring path for all scores. In this paper, we present algorithms for online scoring. Our algorithms are developed as an extension of the recent multi-label classification task. First, we learn the optimal score path through the combination of labels and scores. Second, we provide algorithmically rigorous evaluation results that show that the performance of the algorithms are comparable or better than the existing state-of-the-art algorithms. Experiments using both synthetic and real data show that our algorithms are efficient and robust to a significant loss in accuracy, especially when a novel scoring path is assigned to the scores.