A Fast Convex Relaxation for Efficient Sparse Subspace Clustering – While deep neural networks have made impressive progress in many computer vision applications, they are still suffering from its limitations in particular when the training data is sparse. In this paper, we propose to tackle these limitations by using a convolutional neural network (CNN) to train a CNN for a single sparse subspace clustering problem. Our first model is a convolutional neural network with a convolutional convolutional layer. The CNN is trained with two layers of LSTMs and each layer is used to learn a convolutional convolutional sparse subspace. By combining the learned sparse subspaces, the CNN is trained to learn the corresponding sparse subspace using the training set. Through extensive numerical experiments, we demonstrate the effectiveness of our CNN for solving the sparse subspace clustering problem.

In this work, we propose ToSAR, a deep reinforcement learning (RL) robot that uses its speech recognition capabilities for natural language processing. ToSAR is an automatic saliency-based recurrent agent that learns to distinguish text from images, therefore solving the problem of speech recognition from natural context. ToSAR is trained on real-world data, which involves a speech recognition problem and a human-robot interaction domain. The first approach is a two-stage learning approach that consists of using three different types of reinforcement learning (SRL), namely, learning from input and reinforcement learning, or neural-sensor-sensing, respectively. We design two variants of ToSAR learning module, namely, NeuralNet with a 3D neural network-based approach, and ToSAR that requires a human to be able to recognize input text given a natural context. ToSAR uses reinforcement learning techniques to learn from input and to predict future actions. ToSAR is evaluated on real-world and synthetic data and shows promising results.

Evaluating Neural Networks on ActiveLearning with the Lasso

Hierarchical Multi-View Structured Prediction

A Fast Convex Relaxation for Efficient Sparse Subspace Clustering

An Empirical Study of Neural Relation Graph Construction for Text Detection

Improving Speech Recognition with Neural NetworksIn this work, we propose ToSAR, a deep reinforcement learning (RL) robot that uses its speech recognition capabilities for natural language processing. ToSAR is an automatic saliency-based recurrent agent that learns to distinguish text from images, therefore solving the problem of speech recognition from natural context. ToSAR is trained on real-world data, which involves a speech recognition problem and a human-robot interaction domain. The first approach is a two-stage learning approach that consists of using three different types of reinforcement learning (SRL), namely, learning from input and reinforcement learning, or neural-sensor-sensing, respectively. We design two variants of ToSAR learning module, namely, NeuralNet with a 3D neural network-based approach, and ToSAR that requires a human to be able to recognize input text given a natural context. ToSAR uses reinforcement learning techniques to learn from input and to predict future actions. ToSAR is evaluated on real-world and synthetic data and shows promising results.