Return to Article Details Training autoencoders on their own outputs causes collapse Download Download PDF