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

The article introduces PURe, a Product-Unit Residual Module designed for deep vision networks, which incorporates a 2D Product Unit with a log-domain formulation to facilitate multiplicative local interactions. PURe serves as a drop-in replacement for traditional residual units and has been tested in residual CNNs for image classification and 2D encoder-decoder networks for volumetric CT segmentation. The implementation of PURe shows significant improvements in accuracy and parameter efficiency across benchmarks like Galaxy10 DECaLS, ImageNet, and CIFAR-10, suggesting that multiplicative interactions can enhance the performance of deep residual architectures while using fewer parameters.