PURe's plug-and-play module lets smaller CNNs beat deeper ResNets

Modern vision networks rely heavily on additive local transformations, while explicit multiplicative interactions—despite their theoretical appeal—have been largely ignored due to optimization instability. PURe (Product-Unit Residual Module) tackles this head-on with a 2D product unit built on a real-valued log-domain formulation. This design keeps multiplicative aggregation stable within deep residual hierarchies, allowing PURe to act as a drop-in replacement for native residual units in existing architectures. The researchers from Hochschule Koblenz and the University of Applied Sciences embedded PURe into residual CNNs for image classification and into 2D residual encoder-decoder networks for slice-based volumetric CT segmentation.

Across Galaxy10 DECaLS, ImageNet, and CIFAR-10, PURe consistently outperformed standard residual CNNs and achieved a more favorable accuracy-parameter trade-off. Notably, moderately deep models equipped with PURe could match or surpass substantially deeper ResNet baselines while using far fewer parameters. On the AMOS benchmark, PURe also improved slice-based CT segmentation under 3D case-level evaluation. These results demonstrate that explicit multiplicative local interaction is a practical and effective design primitive for deep residual vision networks, potentially enabling more efficient model architectures for edge deployment and medical imaging tasks.