If too many hidden nodes are used, the autoencoder will essentially memorize the source data -- overfitting the data, and the model won't generalize when previously unseen data is encountered. As a ...