Neural network training loss optimization utilizing the sliding innovation filter