Combining Slow and Fast: Complementary Filtering for Dynamics Learning