Reviews: Professor Forcing: A New Algorithm for Training Recurrent Networks

The idea in this paper is interesting and well motivated. When training an RNN for generation, using the likelihood of the observed data is not the proper criterion. Even though the problem and approach are interesting, I find the description of the training objective (Sec. Precise remarks follow: - The random variable y is present in two expectations in Eq. (1). Given an RNN with a given sequence of inputs, all the y can be computed.