Abstract: We propose using autoencoders to create a data-driven classification of optical galaxy spectra. Autoencoders are unsupervised neural nets that are trained to learn a reduced dimensionality representation of the data. Dimensionality reduction with autoencoders can be thought of as a non-linear generalization of Principal Component Analysis (PCA). We find that, compared to PCA, the autoencoder is better able to reconstruct non-linear features like broad emission lines in QSOs. The autoencoder latent space naturally separates out different classes of galaxies as determined by other criteria like line ratios. Traversing the autoencoder latent space between two spectra produces a series of synthetic spectra that smoothly interpolates between the two. These synthetic spectra look physically plausible with high-level features that vary smoothly.