NeurIPS 2020

Autoencoders that don't overfit towards the Identity

Meta Review

All reviwers have appreciated the theoretical analysis in the paper that highlights that denoising autoencoder can still overfit towards identity. The paper also proposes a way to avoid this. Authors should address the points raised in the reviews about more clarity on experimental setting, giving more intuition on why its important to prevent overfitting towards identity, and commenting on how the intuitions can be carried forward to non-linear autoencoders. Authors should also consider rephrasing the BI statement to something more formal as pointed out in one of the reviews.