Autoencoders are used to reduce the dimensionality of the feature space. They can capture nonlinearities that other dimensionality reduction teqniques like PCA can not. Autoencoders are build by training the model to reproduce the input. In this case you can split the data set into three:
Train you model using the training ...