Undercomplete Autoencoder uses less dimension (number of input units) for latent representation than the input.

The opposite is called Overcomplete Autoencoder.