Partially Linear Autoencoders for Manifold Learning and Dimensionality Reduction
It is demonstrated that imposing a linear encoder preserves most of the representational capacity of the autoencoder, provided the decoder remains nonlinear, and suggested that the nonlinear decoder is the critical component for manifold learning, rather than the encoder.
Louen Pottier, Louis Lesueur, Anders Thorin
· 1 citation