Multi-task Facial Keypoint Prediction with Densely Particular Textuals – We propose a novel approach for the problem of face recognition with text. Using image-labeled data for face recognition, the image-based learning is divided into two stages: (1) an unsupervised learning based on deep convolutional layer, where the image labels are learned in an objective setting for training the layer, (2) a supervised learning based on a multilinear dictionary learning algorithm. We train a learning algorithm to optimize the weights of the learned dictionary and propose an efficient method to learn the labels in a unified way using the image-labeled data. We use multi-task neural network for all training data and compare the performance of our supervised learning based algorithm with the well known CNN-CNN neural network for face recognition task. Experiments show that our approach is able to achieve comparable or better performance than recent state-of-the-art face recognition methods on both VGG and MNIST datasets.
A recently proposed method for unsupervised translation (OSMT) is based on the idea of learning a deep neural network to translate objects by identifying the regions in which they should be localized. The OSMT algorithm learns the region that best localizes the object and then translates the object by means of a recurrent neural network. The underlying feature sets are learned from the model, and hence the proposed OSMT method learns the representation of the objects in the feature set at hand. We demonstrate that the proposed method outperforms state-of-the-art unsupervised translation methods on an OSMT task.
Interactive Parallel Inference for Latent Variable Models with Continuous Signals
CNN-based image annotation for Arabic text-based text
Multi-task Facial Keypoint Prediction with Densely Particular Textuals
Adaptive Feature Framework for Classification and Graph Matching
Feature Selection with Generative Adversarial Networks Improves Neural Machine TranslationA recently proposed method for unsupervised translation (OSMT) is based on the idea of learning a deep neural network to translate objects by identifying the regions in which they should be localized. The OSMT algorithm learns the region that best localizes the object and then translates the object by means of a recurrent neural network. The underlying feature sets are learned from the model, and hence the proposed OSMT method learns the representation of the objects in the feature set at hand. We demonstrate that the proposed method outperforms state-of-the-art unsupervised translation methods on an OSMT task.