End-to-End Action Detection with Dynamic Contextual Mapping – We propose a deep-learning method to predict the action using contextual MAPs for action prediction in real-time. The state of the art works use a mixture of the following two strategies. First the action prediction is used to predict whether two events should be considered as a single event and by what order they should be considered. Second the Mapped Contextual Mapping is used to predict whether the action should be considered as a chain event or a sequence of events. Finally the Contextual Mapping for the action prediction is used to predict the sequence of events from the contextual data in order to predict the action’s outcome. Compared to state-of-the-art deep learning methods, our method outperforms them in terms of accuracy and speed.

We propose an efficient algorithm to explore spatial ordering in a convolutional neural network. The goal is to use the ordered state information from the convolutional layers to determine the ordering of a recurrent neural net to find optimal solutions. We describe a deep neural network architecture in which the goal is to optimize the order of information in each layer to obtain a final solution. Our architecture makes use of the information obtained from prior state information to learn a global context, based on a hidden model of the state, that takes information from the layers as hidden state, and predicts how to perform the search for each hidden state. We present three experiments of four different levels in the Deep Network architecture, where our strategy was to scale to a large number of layers before starting to explore the order of information, in order to minimize the search over all data. We are also able to train a deep net with the same strategy. Hereby we provide an overview of our approach using the knowledge given by the previous layers of the network.

Fast Convolutional Neural Networks via Nonconvex Kernel Normalization

Learning the Top Labels of Short Texts for Spiny Natural Words

End-to-End Action Detection with Dynamic Contextual Mapping

Diving into the unknown: Fast and accurate low-rank regularized stochastic variational inference

Dynamic Metric Learning with Spatial Neural NetworksWe propose an efficient algorithm to explore spatial ordering in a convolutional neural network. The goal is to use the ordered state information from the convolutional layers to determine the ordering of a recurrent neural net to find optimal solutions. We describe a deep neural network architecture in which the goal is to optimize the order of information in each layer to obtain a final solution. Our architecture makes use of the information obtained from prior state information to learn a global context, based on a hidden model of the state, that takes information from the layers as hidden state, and predicts how to perform the search for each hidden state. We present three experiments of four different levels in the Deep Network architecture, where our strategy was to scale to a large number of layers before starting to explore the order of information, in order to minimize the search over all data. We are also able to train a deep net with the same strategy. Hereby we provide an overview of our approach using the knowledge given by the previous layers of the network.