Neural Sequence-to-Sequence Models with Adversarial Priors – In this paper, we propose a novel recurrent neural network (RNN) as a means of learning to answer unsupervised questions. It is particularly well suited to question answering tasks where the answer space is unknown (e.g. answering in a computer). To this we propose an approach to learn a neural network based on a recurrent neural network (RNN) that adaptively learns answer spaces to be more informative to a questioner. We also show that with a suitable non-recurrent layer, we can learn a non-recurrent RNN on the same task. We show that the RNN outperforms the current state-of-the-art RNNs in terms of recall, and achieves better performance. The model also enables us to show that the model can be used in order to learn a sequence-to-sequence model and obtain a better performance. Furthermore, we present the results of our method on a real world example.

We propose a novel framework for unsupervised reinforcement learning by using temporal dependencies between signals. The underlying temporal relationship is a set of continuous variables in which every interaction is modeled as local, i.e., a new variable, and each time step is modeled as an interval, and each interval represents a time dependency of a learned action, the action’s behavior, and the associated dependency. The proposed framework can be used to learn temporally coherent actions, and it is especially useful for learning in environments that exhibit frequent interactions. Experiments on several challenging benchmarks show that our method outperforms both supervised and unsupervised reinforcement learning.

Efficient Bipartite Markov Chain Monte Carlo using Conditional Independence Criterion

An Experimental Evaluation of the Performance of Conditional Random Field Neurons

Neural Sequence-to-Sequence Models with Adversarial Priors

Improving Recurrent Neural Networks with Graphs

Rationalization of Symbolic ActionsWe propose a novel framework for unsupervised reinforcement learning by using temporal dependencies between signals. The underlying temporal relationship is a set of continuous variables in which every interaction is modeled as local, i.e., a new variable, and each time step is modeled as an interval, and each interval represents a time dependency of a learned action, the action’s behavior, and the associated dependency. The proposed framework can be used to learn temporally coherent actions, and it is especially useful for learning in environments that exhibit frequent interactions. Experiments on several challenging benchmarks show that our method outperforms both supervised and unsupervised reinforcement learning.