High-dimensional Asymptotics of Denoising Autoencoders
The episode discusses 'High-dimensional Asymptotics of Denoising Autoencoders,' detailing a paper by Cui and Zdeborová. The hosts explain that the paper provides closed-form equations to predict a denoising autoencoder's performance, showing that adding a skip connection fundamentally makes the network non-linear and superior to simple PCA.