To understand what, how and when these unlikely incidents might occur, the researchers have used an artificial neural network they call a variational autoencoder (VAE).