This paper proposes the class activation value (CAV), which owns similar properties of CAM, while CAV is versatile in various types of inputs and models, and proposes a novel method named CAV Learning (CAVL) that selects the true label by the class with the maximum CAV for model training.