Bringing machine learning into a state estimator (learning the Kalman gain)

Where and how could you incorporate machine learning into a Kalman-filter-based state estimator? In particular, what would it mean to 'learn the Kalman gain', and what are the trade-offs versus a classical filter?