Decomposing learned weights into frequency components to isolate task-relevant circuits from noise or adversarial patterns.