Autoencoder Node Saliency: Selecting Relevant Latent Representations