Export Reviews, Discussions, Author Feedback and Meta-Reviews

Summary: The paper proposes a modification of the traditional bidirectional RNN, which predicts output at time t from forward hidden state t-1 and backward hidden state t 1, without getting information from x_t. This allows training the network to map directly from inputs X to outputs Y. The authors test the capability of this architecture to compute the likelihood of completed gaps in sequences, using GSN and NADE - the results are compared to "Bayesian MCMC" inference in standard RNNs. The paper shows that for longer gaps, GSN using the specified BRNN architecture is computationally cheaper and competitive in quality with RNNs and NADE is inferior in quality to GSN. Quality and Clarity The models and the inference, which was used to compute the NLL scores, are both clearly described.