An assumed property of Neural Network for Computer Vision tasks wherein transformations (rotations, scaling, movement) should not change the output scores by much.