A technique that traces a neural network's prediction back through its layers to identify which input features most influenced the output.