Learning Deep Transform Architectures using Label Class Regularized Deep Convolutional Neural Networks – Deep neural networks (DNNs) have become the standard tool for many tasks, like image recognition and semantic clustering. However, the quality of the results obtained using the DNNs and their performance is often limited due to their high power. In this work, we show that the power-hungry DNNs perform better than others at several tasks. In particular, by using a simple and efficient DNN, we demonstrate that even a small sample of a DNN outperforms the best DNN in performance, and is comparable to the best DNN in performance in state-of-the-art ImageNet benchmark.
This paper presents an implementation of a novel model-based deep learning method that employs a supervised learning framework. To achieve such a task, we build a hierarchical deep neural network that combines supervised learning of an unknown class, a supervised learning process used in the supervised learning process, and an unlabeled model. We show that this approach works well for supervised learning of complex features such as faces, given that the supervised learning involves only a few examples in each feature space. Then, the unlabeled CNN can train to predict the pose for the face in a certain image and infer the pose for each image in the hidden space. The proposed approach outperforms the current state-of-the-art supervised learning methods on two challenging datasets, namely LADER-2007 and MYSIC 2012. The experimental evaluation on both datasets provides promising results.
Unsupervised Video Summarization via Deep Learning
Learning Deep Transform Architectures using Label Class Regularized Deep Convolutional Neural Networks
Word sense disambiguation using the SP theory of intelligence
Structured Highlight Correction with Multi-task OptimizationThis paper presents an implementation of a novel model-based deep learning method that employs a supervised learning framework. To achieve such a task, we build a hierarchical deep neural network that combines supervised learning of an unknown class, a supervised learning process used in the supervised learning process, and an unlabeled model. We show that this approach works well for supervised learning of complex features such as faces, given that the supervised learning involves only a few examples in each feature space. Then, the unlabeled CNN can train to predict the pose for the face in a certain image and infer the pose for each image in the hidden space. The proposed approach outperforms the current state-of-the-art supervised learning methods on two challenging datasets, namely LADER-2007 and MYSIC 2012. The experimental evaluation on both datasets provides promising results.
Leave a Reply