I am looking for more examples of hybrid neural net/Stan models like Maggie Lieu’s excellent keynote at StanCon Helsinki 2018. Going straight to the relevant part:
That talk describes swapping out a 3 ODE deep halo mass function with a TensorFlow implementation that approximates it from simulated data. Works great apparently. She calls it a PDF emulator.
Are there other examples out there? One can imaging defining priors this way in addition to likelihood elements.
At this point I am using Maggie’s example to show that Bayesian modeling and neural nets can play nicely together. And that they bring complimentary but different strengths to bear on interesting problems. This is for a grant but perhaps some research as well.