Improving Representation Learning in Autoencoders via
Multidimensional Interpolation and Dual Regularizations
Abstract
Autoencoders enjoy a remarkable ability to learn
data representations. Research on autoencoders
shows that the effectiveness of data interpolation
can reflect the performance of representation learning. However, existing interpolation methods in
autoencoders do not have enough capability of
traversing a possible region between datapoints on
a data manifold, and the distribution of interpolated
latent representations is not considered. To address
these issues, we aim to fully exert the potential
of data interpolation and further improve representation learning in autoencoders. Specifically, we
propose a multidimensional interpolation approach
to increase the capability of data interpolation by
setting random interpolation coefficients for each
dimension of the latent representations. In addition,
we regularize autoencoders in both the latent and
data spaces, by imposing a prior on the latent
representations in the Maximum Mean Discrepancy (MMD) framework and encouraging generated
datapoints to be realistic in the Generative Adversarial Network (GAN) framework. Compared
to representative models, our proposed approach
has empirically shown that representation learning
exhibits better performance on downstream tasks
on multiple benchmarks