A new way to build neural networks could make AI more understandable