An explainability technique that uses model gradients to identify which input features most influence predictions.