Architecture search without using labels for deep autoencoders employed for anomaly detection

Methods, systems, and computer-readable storage media for defining an autoencoder architecture including a neural network, during training of the autoencoder, recording a loss value at each iteration to provide a plurality of loss values, the autoencoder being trained using a data set that is associated with a domain, and a learning rate to provide a trained autoencoder, calculating a penalty score using at least a portion of the plurality of loss values, the penalty score being based on a loss span penalty PLS, a convergence penalty PC, and a fluctuation penalty PF, comparing the penalty score P to a threshold penalty score to affect a comparison, and selectively employing the trained autoencoder for anomaly detection within the domain based on the comparison.