domain-adversarial training

Using adversarial objectives and gradient-reversal layers to suppress nuisance or subject-specific signals (e.g., speaker identity) so learned representations become invariant to those factors while preserving task-relevant information.