Revisiting Denoising Auto-Encoders

Denoising auto-encoders (DAE)s were proposed as a simple yet powerful way to obtain representations in an unsupervised manner by learning a map that approximates the clean inputs from their corrupted versions. However, the original objective function proposed for DAEs does not guarantee that denoising happens only at the encoding stages. We argue that a better representation can be obtained if the encoder is forced to carry out most of the denoising effort. Here, we propose a simple modification to the DAE's objective function that accomplishes the above goal.