Export Reviews, Discussions, Author Feedback and Meta-Reviews

Neural Information Processing Systems 

The basic idea of this paper is to replace the MCGSM (mixture of conditional Gaussian scale mixtures) of [38] with a version where a continuous-valued hidden state vector h is maintained in a LSTM. This is used as a model of natural images and assessed by a density estimation task (secs 3.2 and 3.3, Tables 1-3), and for texture synthesis and inpainting (sec 3.4). The model for p(x_ij h_ij) (l 150) is in fact not specified at all. Given that h is a continuous-valued vector (es per eq 6) we need to see some functional form. RNADE [41] is designed for fixed-length vectors.