Present paper suggests a new ultrasonic inspection technique to classify the defects occurred by casting of cast steel using the neural network. Reflected echo waves are hard to discriminate the defects by the way of wave observation because of receiving the strong effect of surface roughness. Learning and classification of neural network for the subject of both vacancy defect and gas defect were carried out to the reflected echo. Defects could be classified by specifying the ratio of defect and non-defect as well as the number of learning data.