We present a neural network autoencoder structure that is able to extract the most important latent spectral features from observed spectra and reconstruct a spectrum from those features. Because of the training process, the network is able to reproduce a spectrum of high signal to noise ratio that does not show any spectral peculiarities. Such generated spectra were used to identify various emission features among spectra acquired by multiple surveys using the HERMES spectrograph. Emission features were identified by a direct comparison of the observed and generated spectrum.
