likelihood function

I've spent a lot of time in the last few weeks thinking about the differences between discriminative and generative models. One big difference is that a generative model has a concept (implicit or explicit, depending) of a likelihood function, or probability for the data given some kinds of parameters. I made this explicit today, making the generative model that is being studied by Villar (NYU) and Huang (NYU) and me a likelihood function that includes within its parameter space the true, correct generative likelihood function. We implemented an iterated least-squares optimizer for it and it seems to work and give sensible results. We aren't understanding some of the dependencies we are seeing on hyper-parameters.