Neural networks naturally learn many transformed copies of the same feature, connected by symmetric weights.