Neuroscientists find a way to make object-recognition models perform better