PURe: A Plug-and-Play Product-Unit Residual Module for Vision Networks

A recent study presents PURe, a product-unit residual module aimed at improving deep vision networks through the integration of explicit multiplicative local interactions. This module utilizes a 2D product unit with a real-valued log-domain formulation, effectively tackling the optimization instability that has hindered the application of product units in deep learning architectures. PURe can seamlessly replace standard residual units and has been applied in residual CNNs for image classification as well as in 2D residual encoder-decoder networks for slice-based segmentation of volumetric CT data. Tests conducted on Galaxy10 DECaLS, ImageNet, and CIFAR-10 demonstrate notable enhancements in accuracy and a better accuracy-parameter balance. The paper can be accessed on arXiv under the identifier 2505.04397v3.