Adjusting the magnitude of tensors during computation to keep values in a stable numerical range for low-precision arithmetic.