Learning in high dimensional spaces is computationally expensive because of the curse of dimensionality. Consequently, there is a critical need for methods that can produce good low dimensional representations of the raw data that preserve the significant structure in the data and suppress noise. This can be achieved by an autoencoder network consisting of a recognition network that converts high-dimensional data into low-dimensional codes and a generative network that reconstructs the high dimensional data from its low dimensional codes.Experiments with images of digits and images of faces show that the performance of an autoencoder network can sometimes be improved by using a non-parametric dimensionality reduction method, Stochastic Neighbour Embedding, to regularize the low-dimensional codes in a way that discourages very similar data vectors from having very different codes.