Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for jointly training an encoder neural network and a decoder neural network. In one aspect, a method comprises: updating current values of a set of encoder parameters and current values of a set of decoder parameters using gradients of a reconstruction loss function that measures an error in a reconstruction of multi-modal data from a training example, wherein: the reconstruction loss function comprises a plurality of scaling factors that each scale a respective term in the reconstruction loss function that measures an error in the reconstruction of a corresponding proper subset of feature dimensions of the multi-modal data from the training example.