Laplace approximation now uses Multivariate Normal's

[dustinvtran]

• issues | fixes laplace approximation should use normal as approximating families #329

Previously, ed.Laplace ran MAP and then printed out the precision matrix. That is not useful. Now ed.Laplace properly takes in multivariate normal random variables as input, optimizes their means via MAP, and finally sets their covariance matrices to be the inverse of the observed Fisher information.

See test_laplace.py for the many ways you can instantiate ed.Laplace. They follow the many subclasses of MultivariateNormal random variables.

Remarks

• I replaced the ed.hessian utility function in favor of TensorFlow's tf.hessians. This significantly speeds up their calculation.