Problem :** Compared to discriminative models, generative has limited performance because it is difficult to approximate intractable probabilities with maximum likelihood via a back-prop.
Idea :** Introduce a discriminator to learn adversarially.
objective : Train the discriminator to be good at discriminating between generated and real data, while the generator creates data that the discriminator is not good at discriminating.
baseline : restricted Boltzmann machines(RBM), deep Boltzmann Machines(DBM), deep Belief network(DBN)
data : mnist, Toronto Face Database, CIFAR10
result : Paren window-based log-likelihood estimates based on SOTA
contribution : Unlike RBM, it does not use markov chains, but only gradients to learn.
Limitations or things I don’t understand: 4.2. Convergence of Algorithm 1 doesn’t make sense to me
