Surround the Nonlinearity: Inserting Foldable Convolutional Autoencoders to Reduce Activation Footprint | IEEE Conference Publication | IEEE Xplore